The Briefing Desk

Baidu published ERNIE-5.5-Flash-Base on Hugging Face with a licence that gives its release date as 5 November, so it predates this week's 7–14 November coverage window. It's a sparse mixture-of-experts language model with 280B total parameters, 14B of them active per token, trained from scratch on 33.5T tokens. Unusually, Baidu released the checkpoint from the end of pre-training, before any mid-training or post-training. That makes it a base model for people who want to do their own instruction tuning, domain adaptation or RL, not a chatbot.

Most layers (37 of 43) use "Heavily Compressed Attention", which caches one entry per 128 tokens, and 6 layers keep full-context latent attention with sparse attention. Baidu also uses several independent routers per MoE layer instead of one. A component called ECHO reuses the final layer for multi-token prediction, and the same layer doubles as a built-in draft model for speculative decoding. Training context was 128K, and the released config enables 1M tokens for fine-tuning and inference. Every benchmark (for example MMLU-Pro 78.4 versus 69.2 for DeepSeek-V4-Flash-Base) is Baidu's own evaluation, not independently reproduced. The licence is Baidu's own ERNIE 5.5 Flash Base Model License Agreement, not a standard open-source licence, and it grants rights only within a "Territory" that excludes the European Union, the United Kingdom and South Korea. The release is text-only.

Why it matters. Base checkpoints at this scale are rare, and most open releases arrive already tuned to the vendor's preferences. This one lets teams choose their own post-training recipe.

Google made Gemini 4 Carbon generally available on 10 November. Two weeks earlier it had been released only to vetted cybersecurity defenders through the Fairwind Program. Google positions Carbon above Gemini 4 Argon, which it still recommends for most software engineering and knowledge work. Carbon is for the hard tail: multi-step coding and debugging, vulnerability work and open-ended research. Google says Carbon spends more tokens than Argon to explore, reason and run verification loops.

Paying developers can call it through the Gemini API in Google AI Studio and Android Studio, and enterprises through Gemini for business. Until 30 November it costs $2 per million input tokens and $10 per million output tokens, the same introductory rate as Argon. After that the price rises to $4 and $20. Google AI Ultra subscribers also get it in Antigravity, the Gemini app and AI Mode in Search. Alongside the model, Google is previewing /swarm in Antigravity, which splits one prompt across a coordinated team of agents: Carbon for planning and auditing, lighter models for parallel work, and adversarial verifiers that test each change.

Google claims first place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at "more than 2X lower cost per task" than the next external model. These are third-party benchmarks, but the rankings and cost comparisons are Google's own reporting. The strongest cyber results (SEC-bench Pro, ExploitGym, CWE-bench) apply to "Carbon for cyber" through Fairwind, not to the public model. On safety, Google reports 92% on its own 375-task agentic honesty benchmark, up from 70% for Argon. It also acknowledges that prompt-injection robustness "remains an area for continued improvement."

Why it matters. Carbon is a model that deliberately trades more tokens per task for higher accuracy on long jobs. For teams running agents, the real question is cost per finished task, not price per token, and Google is pricing Carbon to win that comparison during November.

Meta has released Muse Ember (muse-ember-1.0), the smallest model in its Muse family, on its Model API and in Muse Code. Meta positions it as an execution model, not a planner: a larger Muse Wonder or Muse Spark model breaks the work down, and Ember handles the volume, including sub-agent coding tasks, terminal commands, tool calls, request routing and batch extraction across thousands of files or images.

The specs are generous for a small model. Ember accepts text, image, video, audio and PDF input, has a 1,048,576-token context window and can emit up to 524,288 output tokens. Pricing is $0.10 per million input tokens and $0.50 per million output tokens for requests under 200K input tokens, rising to $0.30 and $1.50 above that, with cached input at $0.004. It is served through Responses, OpenAI-compatible Chat Completions and Anthropic-compatible Messages endpoints, and Meta says it accepts the same request body as Muse Spark, so switching is mostly a model-ID change.

One detail is easy to miss: reasoning effort defaults to "max". Meta has not published independent evidence for how close Ember gets to its larger models, so its performance claims remain vendor-reported.

Why it matters. Planner-worker splits are becoming the standard shape of production agents, and the economics depend on the worker being cheap. At these prices, fanning out dozens of parallel sub-tasks becomes realistic, provided the model holds up on your tasks.

Meituan has published the weights of LongCat-2.5 on Hugging Face under the MIT License, a week after introducing the model. It is a natively multimodal mixture-of-experts model with 1.6 trillion total parameters, about 48 billion active per token, aimed at long-horizon agentic coding, full-project builds and visual tasks such as turning screenshots into web pages.

The model card is more interesting for its infrastructure notes than its leaderboard. Meituan says LongCat-2.5 was trained and served on domestic Chinese AI accelerators, and that it applies W4A4 quantisation to expert weights and activations so they run natively in INT4 on that hardware. Meituan reports that this setup gives about 43% higher prefill and decode throughput than an 8-bit baseline, and that its RL pipeline keeps token log-probabilities bitwise-identical across training, prefill and decoding, a known source of instability in reinforcement learning on large models.

On benchmarks, Meituan's own table puts LongCat-2.5 close to leading closed models on several agentic tests, but it is mixed: it trails on Terminal Bench 4.0 and SWE Atlas test writing, for example. Many scores are measured in-house, and the card itself calls those indicative only.

Why it matters. A frontier-scale multimodal agent model under a plain MIT licence is unusual, and so is open documentation of a training and INT4 serving stack built outside NVIDIA's ecosystem. It's a concrete data point on how far Chinese labs have moved onto domestic hardware.

MiniMax released H3-Image on 11 November, an image generation and editing model built on its H3 family. The weights are on Hugging Face and it is also available through MiniMax's API. One checkpoint handles both text-to-image and editing with up to 16 reference images. It generates about 16 megapixels natively, from 4096×4096 to 6272×2688, with no upscaler.

The model card is unusually specific. It's a dense, 33.1-billion-parameter, single-stream diffusion transformer, the H3 architecture with the audio branch removed. It reads layer-50 hidden states from MiniMax-H3's Qwen3-VL text encoder and uses a VAE with 16× spatial downsampling, so each transformer token covers a 32×32 pixel patch. That coarse patching is a large part of why 16 MP is affordable. The model is distilled to six sampling steps with no classifier-free guidance. MiniMax claims under 10 seconds per 4K image and "up to 5×" the speed of mainstream alternatives. Those speed and quality claims are the vendor's and haven't been independently measured yet.

"Open" needs reading carefully. The licence is MiniMax's own H3-Image Community License, not an OSI licence. It grants rights only outside its "Excluded Territories", which are the European Union, the United Kingdom, South Korea and the United States, and it forbids using the weights or their outputs there. Elsewhere, companies with more than $1 million in annual revenue, or anyone offering a hosted service, need separate written authorization from MiniMax.

Why it matters. Native 4K output in a handful of steps changes what's practical for print, product and design work, where upscaler artefacts have been the giveaway. Where the licence applies, an open-weight model at this level also gives self-hosters a credible alternative to closed image APIs.

ByteDance's Seed team introduced Seedream 5.5, its next image generation model, on 9 November. The headline feature is stable multi-turn editing: Seed says the model preserves unedited regions at the pixel level without colour shifts, and demonstrates 100 successive edits on one portrait while still changing small objects such as rings and eyeglass frames. It also generates native RGBA images with full and partial transparency, so elements can go straight into layouts without background removal. Seed reports more natural skin, lighting and small-text rendering, with a severe text corruption rate of 0.4% in its tests.

The evaluation is entirely ByteDance's own. In production use-case ratings on its MagicArena platform, usability in film and TV and in office work was more than double Seedream 5.0 Pro's, and in preference comparisons Seed says the model leads GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure. None of this has been independently checked. Seed's Chinese-language announcement says it is rolling out gradually and lists four ways to try it: Jimeng on the web (Create, then Generate or Canvas), image generation in the Doubao app and website, Doubao's work skills, and the Volcengine Ark experience centre, where it appears as Doubao-Seedream-5.5. A Pro version is promised, without a date.

Why it matters. Drift across repeated edits is one of the main reasons generated images still need manual cleanup. A model that leaves untouched pixels alone makes iterative, conversational editing usable for production work, and native transparency removes a separate cut-out step.

ElevenLabs released two speech-to-text models on 10 November. Scribe v3 is for recorded audio. ElevenLabs says it makes 27% fewer word errors than Scribe v2 across 102 languages and 17% fewer speaker-labelling (diarization) errors on recordings with three or more speakers. Scribe v3 Live is the low-latency realtime variant. ElevenLabs says it makes 43% fewer errors than Scribe v2 Realtime, measured as semantic word error rate, and adds diarization to realtime transcription for the first time.

For voice agents, the more interesting change is turn detection. Instead of waiting out a fixed silence, Scribe v3 Live uses conversational context as well as audio to predict when a speaker has finished, so a caller who pauses halfway through a number isn't cut off. The turn detection ElevenAgents already used is now available in the API. Other controls include keyterm prompting with up to 1,000 terms, transcript formatting rules written in plain English, a choice of digits or words for numbers, and, in Live, a multilingual mode for speakers who switch languages mid-sentence.

All figures come from ElevenLabs' own evaluations. Multilingual accuracy was measured on an internal version of the FLEURS benchmark, and diarization on an in-house dataset of real meetings, agent calls and voice notes that the company says it will open-source. Dutch energy supplier Eneco reports that its voice agent recognised postcodes correctly on the first try 40% more often in a pre-launch test.

Why it matters. The transcription errors that hurt products are wrong names, numbers and speakers, and agents that interrupt callers. Moving turn prediction into the transcription model follows the same direction as several full-duplex voice launches this week.

Anthropic opened a private beta of self-hosted Claude Managed Agents on 12 November. It is available only through Anthropic account teams: there is no public sign-up or self-serve download. Managed Agents is Anthropic's set of APIs for long-running autonomous agents: the agent harness, sandboxes, credential handling and session history. Until now Anthropic ran that service, and customers could at most host the sandboxes where an agent's tools execute. The self-hosted option moves the rest in-house too: the API server (cma-server), the agent loop, and the storage for agents, sessions and session history, in the customer's own cloud account or data centre. Developers keep using the same API, SDK and CLI.

The documentation is clear about what does not move. The model still runs at Anthropic or at a provider the customer chooses: Amazon Bedrock, Claude models on Google Cloud, Microsoft Foundry or the customer's own gateway. For every model call, the server sends the conversation so far, including tool inputs and results, to that provider. Customers also take on the operating work: running and securing the server, a PostgreSQL 16 or later database (without one, data lives in memory and a restart erases it), and the worker sandboxes, and installing each release themselves. The docs warn that the agent's bash tool is a real shell with the worker's permissions, so isolation depends on running workers in disposable containers or VMs, and each trust level needs its own deployment. Some features exist only on the hosted service. The server software comes under Anthropic's own licence agreement, and access is requested through an Anthropic account team.

Why it matters. Agents that fix builds and triage alerts need standing access to internal code and systems, and many security teams won't let that control plane live in a vendor's cloud. Self-hosting settles where agent state and credentials live. It does not change where inference happens.

AWS made the Web Search API on Amazon Bedrock AgentCore generally available on November 10. Web search was already offered as an MCP tool through AgentCore Gateway; the new InvokeWebSearch API lets any application, custom agent loop or data pipeline call it directly, with no gateway or other resources to set up.

Results come from an Amazon-operated web index. Each one includes a text extract and, when available, the title, URL and publication time. Requests can be filtered by domain and publication date. Calls go through the AWS SDKs or CLI, are authorised with IAM, can stay inside a VPC through PrivateLink, and can be logged as CloudTrail data events. Pricing is $7 per 1,000 queries, the same as the Gateway route but without gateway charges.

AWS hasn't published details on the index's coverage or freshness, so how it compares with established search APIs is still unknown.

Why it matters. Grounding is now a commodity line item inside the major clouds. For teams with compliance requirements, a search call that stays inside their VPC and audit trail may matter more than marginal differences in result quality.

Google DeepMind launched the Gemini Decisions API on 11 November, a new Gemini API endpoint for closed-set judgments, built on a dedicated model called Gemini Decision 1 (gemini-decision-1). It doesn't generate text. It scores candidate answers in a single prefill pass and returns calibrated probabilities. That makes it a classifier and router, not a chat model.

One request can ask up to 1,000 independent questions about a shared input that can mix text, images, audio, video and PDFs. The 262,144-token limit covers the shared input plus all of the questions. There are three question types. Yes/no checks (Google calls them noul) return a probability from 0.01 to 0.99. Choice questions pick from 2 to 255 options and return per-option probabilities. Rubric questions return a weighted score across ordered levels. Every answer comes with a confidence value. Google quotes 150–250 ms end-to-end latency for inputs up to 16K tokens and says it stays nearly flat from 1 to 100 questions; with larger files or the full 1,000-question limit, measure it yourself. Pricing is $0.10 per million input tokens, with output billed at zero, and there is a free tier in AI Studio for prototyping.

The model card shows a narrow model. It has no function calling, structured outputs, search grounding, Batch API or context caching. Google's launch benchmarks, as VentureBeat reported, put it ahead on five of seven shared tests against OpenAI's GPT-6 Luna decision model and TypeSafe's Jev. Jev still led on Mind2Web Multimodal and When2Call. Those are Google's own numbers. On the independent Artificial Analysis Decision Index v2.1, VentureBeat reports, Gemini Decision 1 ranked ninth of 18 at launch with 65.98, just behind Jev 1.14 at 66.23 and well behind Nace AI's Drex Max v1.1 at 72.62. On that suite's all-or-nothing document test, it got every answer right on 33% of documents, against 35.5% for Jev and 63% for Drex Max. Google is entering a category that already has rivals: OpenAI's decision model costs the same per input token, and Simon Willison notes that Jev remains cheaper at 4.2 cents per million.

Why it matters. Many production LLM calls exist only to get one label: which tool to call, whether an input breaks a policy, which model should handle a request. Generating text and then parsing it adds latency, cost and brittle parsing code. A typed probability with a confidence score lets code branch directly and send uncertain cases to a person. Multimodal input also opens up checks on recordings, scans and video that text-only decision models can't do.

Perplexity released Decider v1.3 and Decider Flash v1.3 on 12 November. They are not chat models: they read an input, which Perplexity calls the "state", and return probabilities for questions about it, such as yes/no, a choice among up to 255 labels, or an ordered score. Applications use the numbers for routing, moderation, filtering, reranking and output evaluation, where a threshold is more useful than prose. The larger model is built on Qwen3.8-27B and the Flash model, distilled from it, on Qwen3.5-4B.

The new feature is locate, which extracts verbatim text spans with start and end character offsets and a confidence score, for named-entity recognition, PII detection and extractive question answering. If nothing matches, it returns an empty list rather than inventing an answer. Perplexity's implementation detail is worth noting: locate runs through a separate adapter that loads on the first locate request, rather than changing the classification path. The cost is latency: a request mixing locate and decision questions needs separate forward passes for each type.

Pricing on Perplexity's Decisions API is unchanged from v1.2 at $0.02 per million input tokens for the large model and $0.01 for Flash, with output tokens free. The pplx-decider-latest alias now points to v1.3, and Perplexity warns that probabilities change between versions. Accuracy claims are the company's own; as Unite.AI notes, part of the evaluation compares the model against labels produced or relabelled by frontier LLMs rather than human ground truth.

Why it matters. The release arrives a day after Google launched its own Decisions API, a sign that "small model that returns calibrated decisions" is becoming its own product category. For many production pipelines, a cheap classifier that returns a probability beats calling a frontier model and parsing its text.

Adventists AI released Sophon-Realtime v1 under Apache-2.0 on 8 November and has shipped updates almost daily since, up to v1.3.0 on 14 November. It is a family of full-duplex speech-to-speech models. Full duplex means it listens while you talk, starts drafting while you're still speaking, stops when you interrupt, and ignores a plain "mm-hm". All three builds share one 32B "thinker" that the model card identifies as DuplexJev-32B-Para v1.5. Tag uses an expressive tag-to-speech voice, Flash is a faster and lighter voice, and E2E replaces the speech-token model with a compact talker of about 350M parameters. E2E left research preview on 11 November and now runs as a full build on one GPU.

The thinker writes its reply with inline sound tags like [laughter] or [sigh], and the voice performs them instead of reading them aloud. The vendor's own measurements on one H100 are about 0.6 s to first audio for Tag, 0.31 s for Flash and about 0.25 s for E2E, up to 16 concurrent users per GPU for Tag and E2E and 40 for Flash, with queueing starting at 24 users on Tag. All benchmarks are self-reported, including the "strongest open full-duplex brain" claim (VoiceBench 82.8). The card admits the voice delivers many tags only some of the time, for example a sigh just 25% of the time it's written. The first server start takes 15–20 minutes while kernels and TensorRT engines are built.

Why it matters. Open, self-hostable full-duplex voice with turn-taking and barge-in has mostly been available only from closed APIs. This release also includes a "Live mode" gateway that the vendor says is compatible with the OpenAI GPT-Live interface, so existing clients would only need a new URL and key.

Fish Audio has released Live 3, a full-duplex voice model for real-time voice agents. It keeps listening while it speaks, so a caller can say "mm-hmm", pause mid-sentence, or change their mind without the agent talking over them or cutting in too early.

The pitch is that speed and timing are separate problems. Most voice agents stitch speech recognition, reasoning, speech generation and turn-taking together as pipeline stages, with turn-taking governed by silence thresholds and hand-tuned rules. Fish Audio says Live 3 moves conversational timing into the speech model itself, trained on real human conversations to learn when to acknowledge, wait, respond or stop. "The fix so far has been more rules and more thresholds," said co-founder and CEO Rissa Cao.

The launch post walks through three behaviours. When a caller reads out a phone number in chunks, Live 3 may say "okay, 415" between groups and wait for the rest. It treats "yeah" as a backchannel and keeps talking, but stops when the caller says "Wait, actually, I just need Thursday." And for an order like "two pizzas... no, three," it waits through the corrections instead of logging the wrong item. Fish Audio says the model can decide to respond in as little as 186 milliseconds for short reactions such as "mm-hmm" or quick confirmations, while the full answer is prepared in parallel.

That number needs context. It is a floor for small acknowledgements, not time to first audio for a complete reply, and Fish Audio has published no latency distribution, no turn-taking benchmark scores and no independent evaluation. The evidence so far is three demo excerpts from Fish Audio's own test calls. Live 3 is English-only at launch and is not open-weight.

Live 3 runs on Fish Audio's Agent Platform, through its API, and on-premises for enterprise customers. The Live 3 page lists the agent plan, which includes Live 3, at $0.035 per minute, discounted 50% from $0.07. Phone calls add $0.01 per minute, a dedicated number costs $1.20 per month, and new accounts get $5 of credit without a card. A separate reasoning layer still carries the agent's instructions and tools; premium LLMs are billed at their own token rates.

Why it matters. Turn-taking is where most voice agents still feel robotic: they interrupt during a pause, or talk straight through a correction. Moving that decision into a model trained on real conversation is a credible direction, and Live 3 joins a growing set of full-duplex systems competing on how a call feels rather than just how fast it starts. Until Fish Audio publishes measurements, though, the claim rests on demos.

OpenAI launched two security products on 12 November under its Daybreak programme. Daybreak Red Team puts a coordinated team of agents to work testing web applications. The agents use source code and app context to investigate attack paths and confirm vulnerabilities, validate findings from code scanners, bug bounty submissions and disclosure reports, and re-test whether a patch actually stops an exploit. OpenAI says scope is enforced with isolated environments and Guardian review, and customers control which apps and endpoints the agents can reach. SecurityWeek reports that OpenAI is accepting early-access requests for Red Team.

Daybreak Defense Factory works on the other side of the loop. Agents trace vulnerabilities through a codebase and prepare patches, taking findings from Red Team or from existing security tools, and can scan new commits or run scheduled checks. OpenAI's setup guide describes it as a Preview plugin for ChatGPT workspaces that already have Daybreak access: it scans connected GitHub repositories in Codex cloud and can import findings from Wiz, Snyk and Endor Labs. Engineers review and test proposed patches before merge and can refine them with Codex, and SecurityWeek reports that patches arrive as draft GitHub pull requests and are never deployed automatically. According to The Register, the agents run on OpenAI's latest cyber model, currently GPT-6 Cyber. OpenAI's product pages publish no pricing, benchmark results or false-positive rates, so how well the agents perform on real codebases has not yet been measured in public.

Why it matters. Last week GPT-6 Cyber arrived as a restricted model for vetted researchers. This week it is packaged into products that sit inside a company's development workflow, which moves AI-driven vulnerability discovery and patching from research into routine security operations. The human review step before merge is the control that matters.

OpenAI made inference residency for GPT-6 Luna generally available in India, Japan and Australia on 10 November. For supported requests, both request processing and model inference now take place in the selected country, not only data storage. It covers the Responses and Chat Completions APIs with text and image input and text output. To use it, an organisation needs approved, active Zero Data Retention, then creates a regional API project in the chosen country with Storage & Processing selected.

The Hindu reports that the India option extends the data-residency offering OpenAI introduced there in May 2025, which covered storage only, and that customers send traffic to a corresponding India regional endpoint. The launch covers one model; OpenAI's release note lists no others.

Why it matters. In regulated sectors such as banking and healthcare, where the computation happens can matter as much as where data is stored. Moving inference in-country removes a common blocker to deploying hosted models in these markets.

Notion published a security advisory on 12 November about a Fireworks AI incident that, it says, affected fewer than 1% of Notion workspaces. The investigation is still going. If you have not heard from Notion, the company says it has not identified your workspace as affected and no action is needed right now.

Fireworks first notified Notion on 2 November about a small dataset that, Notion says, was synthetic testing data with no customer data. Notion disconnected the next morning and deleted its data at Fireworks on 4 November. On 6 November Fireworks said malicious actors may have accessed potential customer data; on 8 November it confirmed a final impacted dataset, which Notion says included Notion AI chat transcripts used for the Insights feature. For the vast majority of impacted workspaces, those transcripts were unattributable — not labelled with a user or workspace, and without tool-call results. Notion contacted affected customers on 9 November, disabled Insights, and says it is no longer serving open-weight models through Fireworks, that no other Notion features were affected, and that this was not a breach of its own systems. On 11 November it revoked exposed Notion API keys found in those messages, and on 12 November reported exposed third-party keys to the vendors that issued them. Fireworks, in a bulletin last updated on 9 November, says that after a security incident at a third-party cloud provider on 23 October, an unauthorised party reached its development machines on 24 October and caused additional customer data impact, and that its investigation is ongoing. TechCrunch reports Notion's disclosure.

Why it matters. A vendor used for classification and open-weight inference sat outside the customer's own perimeter, and chat text used for a usage-insights feature was in the dataset later confirmed as impacted. For teams that paste credentials or internal documents into a hosted assistant, the practical question is which jobs actually persist data.

Alibaba tagged Qwen Code v1.0.0-preview on GitHub early on 11 November. It's the first 1.0 build of its Apache-2.0 terminal coding agent. The release notes say it combines the agent capabilities of Qoder, Alibaba's commercial coding product, with Qwen Code for code understanding, implementation, debugging and testing. It works with Qwen and Qoder accounts, multiple model providers and custom compatible endpoints. It also ships as native packages for macOS, Linux and Windows, so you no longer need a separate Node.js install. Session resumption and non-interactive execution are included for scripting and automation.

The preview sits on a separate branch while the 0.27 line continues; v0.27.2 shipped the day before. Alibaba warns that interfaces, configuration options and behaviour may change during the preview. It has since tagged v1.0.0-preview.1 on 12 November and the stable v0.27.3 on 13 November.

Why it matters. Vendors are starting to open-source the agent harness, not just the model. A permissively licensed, embeddable harness that already takes custom endpoints is a realistic base for teams who want their own coding agent without being locked to one provider.

Huawei's Computing Systems Lab in Switzerland has released MACaQ (Model Agnostic Calibrated Quantization) under Apache 2.0, along with a paper on arXiv. It quantizes models one linear layer at a time, using channel balancing, small 32-point Hadamard transforms and bit-flip weight refinement. Every transform stays inside the layer, and that is why the same recipe applies to language, diffusion, speech and 3D-reconstruction models.

The repository has four backends: llama.cpp for 2-, 3- and 4-bit GGUF text models, vLLM for W4A4 or weight-only serving, Diffusers for W4A4 image, video and audio generation, and PyTorch for speech. Huawei says MACaQ can quantize a 122B-parameter model in 64 minutes using 32.5 GB of device memory. On Qwen3.8-27B in vLLM, it reports 2.33× faster prefill and 2.58× faster decode than BF16, with perplexity moving from 2.8463 to 2.8994. A companion tool, MACaQ-R, refines existing GGUF files while keeping them compatible with standard llama.cpp. On Qwen3.6-35B-A3B, Huawei reports lower KL divergence than Unsloth's Q4_K_XL at the same 22.36 GB file size.

All of these are the authors' own measurements. The W4A4 path relies on MXFP4 kernels, so the speed gains depend on hardware that runs 4-bit activations natively. The full MACaQ llama.cpp workflow builds a bundled CUDA runtime instead of stock llama.cpp. Huawei's Ascend kernels are still in upstream pull requests, and full-model Ascend execution still requires runtime integration.

Why it matters. Quantization tools are usually specialised, with one for GGUF text models and others for diffusion or speech. A single calibrated recipe that covers both weights and activations across modalities, with published code, is useful to anyone squeezing models onto smaller hardware. MACaQ-R is the most immediately useful part, because it improves files people already run.

Zyphra released ZONOS3-Mini on November 9, a zero-shot voice-cloning text-to-speech model with open weights under Apache 2.0. It is a sparse mixture-of-experts model with 800 million total parameters and about 100 million active per token, small enough that Zyphra ships runtimes for in-browser WebGPU, Apple Silicon (MLX), ggml, and servers via SGLang and vLLM.

The interesting engineering is in the audio representation. Its predecessor, ZONOS2, used 8 billion parameters on top of the Descript Audio Codec at 86 frames per second, which made real-time generation expensive. ZONOS3-Mini instead predicts tokens from Zyphra's new ZAC-1 codec, which runs at 15 frames per second with ten 4,096-entry codebooks: one semantic codebook distilled from Wav2Vec2-BERT and nine acoustic ones. A causal iSTFT decoder renders each frame to 44.1 kHz audio as soon as it is generated. The 11-layer backbone mixes Gated DeltaNet linear-recurrence layers with global attention in a 3:1 ratio. Zyphra says it trained the model on 112 H100s in 2.5 days.

Zyphra reports under 570 ms to first audio in a browser on a Pixel 9 and under 150 ms on an M3 MacBook Air. These are the company's own figures, not independent measurements, so expect variation by device.

Why it matters. Permissively licensed voice cloning that runs client-side removes both per-minute API fees and the need to send voice samples to a server. That is a real privacy and cost shift for games, narration and voice interfaces. It also lowers the barrier to voice impersonation, which the open licence does nothing to prevent.

The Institute of Foundation Models (IFM), the lab set up by Abu Dhabi's MBZUAI, released PAN 2 on 10 November. You give it a starting scene and an action in words, and it generates video of what happens next. It ships under Apache 2.0 with the model weights, the inference code and, unusually, the pretraining and fine-tuning code.

PAN 2 works in two stages. A language model first writes the next world state as a time-stamped description of events, actor-bound actions and camera motion. Then a 29.2-billion-parameter diffusion transformer, trained from scratch on real-world video chosen for physical dynamics, turns that description into 704×1248 video at 24 fps, up to 20 seconds long. Qwen3.5-9B serves as the frozen text encoder and the Wan 2.2 VAE renders the frames. IFM reports third place overall on the Physics-IQ Verified leaderboard and first among open-weight models. In its own six-dimension evaluation it placed just behind MiniMax-H3, and it completed more multi-step plans than any other model tested, although its model card notes that about two in three simulated plans still contain at least one step that is not carried out. Those comparisons are self-reported. IFM also says long-term world consistency remains unsolved. The code release has a gap too: the reward-learning and model-soup stages that produced the released weights aren't included, so the public code can't fully reproduce the checkpoint.

Why it matters. World models are how robotics and agent teams want to test actions before taking them, and most strong ones are closed. A permissively licensed one with real training code gives researchers something they can actually retrain and adapt, not just sample from.

Google DeepMind has introduced Gemini Robotics 2.5, the next version of its vision-language-action model, which turns camera input and language instructions into motor control. DeepMind tested it on more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. Compared with its previous baseline, it reports a 30% to 100% higher success rate depending on the task, with tasks finished 20% to 30% faster.

The headline result comes from outside the lab. In a proof of concept with Apptronik and Mercedes-Benz, Apollo 2 humanoids running the model sorted automotive parts from bulk bins into ordered trays on production lines in Kecskemét, Hungary, and Sindelfingen, Germany. DeepMind reports a 96% success rate. DeepMind credits much of the progress to using its Gemini 4 Argon model to speed up robotics research. The model works across robot types, from Apollo 2's five-finger hand to the two-gripper Franka Duo.

On safety, DeepMind says the model uses inference-time reasoning to steer around hazards and nearby people. It is also releasing ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests for robots. All performance figures are DeepMind's own, and the factory results come from a pilot, not a production deployment.

Why it matters. Dexterity with deformable and delicate objects, and reliability in real workplaces, have been the main obstacles between lab demos and commercial robots. A 96% kitting rate on live lines is a meaningful data point, though the remaining failures would still be costly at factory scale. The adversarial-safety benchmark is arguably the more lasting contribution, because jailbreaks matter far more when a model controls a physical arm.

Google DeepMind has released AlphaGenome 2, a DNA sequence model that predicts how genetic variants affect gene regulation in specific human cell types and states, not tissue averages. It was trained on more than 100 million individual cell profiles across 68 tissues, drawing on resources such as the Human Cell Atlas. DeepMind says it finds nearly twice as many variants that change gene expression in cells as its previous model.

The technically notable feature is expression conditioning. Researchers can supply their own single-cell RNA-seq data, and the model adapts to cell states it hasn't seen, such as disease progression or drug treatment, without retraining. It works on 1 Mb sequence windows, predicts readouts such as gene expression, chromatin accessibility and histone marks, and adds DNA methylation. It can also impute modalities that weren't measured. Alongside it, DeepMind offers AlphaGenome Atlas, precomputed variant-effect scores across 9 billion human variants.

Access comes with terms. AlphaGenome 2 is free via the AlphaGenome API for non-commercial research only, and commercial research is in private preview. The terms forbid using outputs to train other models. DeepMind states it has not been validated for clinical or diagnostic use. The GitHub client also notes that live predictions suit small-scale analyses rather than genome-wide runs.

Why it matters. Most disease-linked variants sit in non-coding DNA, and their effects often show up only in particular cells. Cell-level predictions help researchers decide which variants and cell types to test experimentally. They point experiments in a direction rather than replace them.

Lyra 2.0 is not this week's news, but it is worth knowing. NVIDIA researchers posted the paper on arXiv on 14 April 2026 and released the model weights and inference code on 15 April. A four-step distillation LoRA followed on 17 April, and the GUI and training code arrived in July, the most recent update (15 July on the project page; the GitHub repo lists 20 July).

The framework generates persistent, explorable 3D scenes from a single image in two stages. A 14-billion-parameter video model built on Wan-14B generates a long camera walkthrough along a trajectory the user chooses. The frames are then reconstructed into explicit 3D Gaussians or meshes that can be rendered in real time or exported to physics engines. NVIDIA's project page shows scenes exported to Isaac Sim for robot navigation, and the GUI lets users plan camera paths to revisit explored areas or move into new ones.

The paper targets the two ways long video generation usually breaks down. Spatial forgetting happens when earlier regions fall out of the model's context and get hallucinated differently on a revisit; Lyra 2.0 keeps per-frame 3D geometry and uses it only to retrieve the most relevant past frames and establish correspondences, leaving appearance to the generative model. Temporal drifting is the slow build-up of small errors; the model is trained on its own degraded outputs so it learns to correct drift instead of propagating it.

NVIDIA says Lyra 2.0 sets a new state of the art in single-image 3D scene generation. That claim rests on its own evaluations (PSNR, SSIM and LPIPS plus qualitative comparisons), and the model card warns that training relies on synthetic data, which may limit how well it handles scenes unlike its training set.

Why it matters. Building simulation-ready 3D environments from one picture could make robot-training worlds far cheaper to produce than scanning real spaces. The core idea, using 3D memory only to route information while a generative model handles appearance, is reusable for any long-horizon world model.

A benchmark posted to arXiv on 11 November tests what most tool-use benchmarks skip: conversation. ToolTurns contains 2,620 problems spanning 1,827 tools and 11,004 API endpoints in 47 categories. Each dialogue opens with a deliberately underspecified request and continues with two context-dependent follow-ups, so an agent has to ask clarifying questions, call tools and remember what was said. The authors also release an open-source user simulator that reveals withheld context through dialogue, with a checker–fixer pipeline to improve its turns.

Across nine language models, the best agent resolved 69.6% of individual requests but completed only 39.8% of full dialogues. Those are two different measures, success on single requests and completion of whole conversations, not a decline over time. Every model scored lower on follow-up requests than on the first one, and rubric scores for dialogue quality ranked models differently from task success. The results come from the authors' own protocol and simulator, so the absolute numbers depend on how closely the simulated user behaves like a real one.

Why it matters. Single-turn tool benchmarks may overstate how well agents hold up in conversations where later requests depend on earlier ones. In this benchmark every model did worse on follow-ups than on the opening request, and completing a whole dialogue was a far harder bar than resolving one request.

Research agents score well on clean benchmark tasks, but researchers usually ask them for one step in a project that is already under way. ARE-bench, posted to arXiv on 10 November, builds 75 tasks from real agent sessions and full workspaces taken from recruited researchers' own projects. The workspaces are replayed without cleanup, so unfinished files, outdated versions and failed attempts are all still there. The tasks span seven research stages and 1,173 binary checklist items.

Across seven frontier models, the strongest passed every checklist item on only 66.2% of tasks, and design and execution ranked among the hardest stages for every model. The key finding is about evaluation itself: three models passed about 69% of tasks when only delivery and general norms were checked, but just 28.4% to 40.0% once requirements carried over from earlier in the session and from the workspace were checked too.

Why it matters. Benchmarks that ignore project history can overstate research agents and hide the differences between them. Respecting earlier decisions and existing files is exactly what users notice when an agent breaks their project.

JoyFrontBench, posted to arXiv on 11 November, separates partial implementation from application success in front-end coding. Each of its 100 problems, covering 764 tasks and 2,596 atomic criteria, comes with an acceptance contract written before generation. A core task passes only if all its mandatory criteria are met, and an application counts as a success only if every core task passes on the same artifact in a single run. The authors evaluated five coding models under two agent harnesses, OpenHands and OpenCode, on a 70-problem test set, keeping the other 30 problems public for validation.

Under OpenHands, the mean average task score across models was 87.69%, while the mean application success rate was 51.52%. Model rankings changed between harnesses, and application success varied across repeated runs.

Why it matters. A high average score can hide an app whose checkout or login doesn't work. For front-end agents, the all-or-nothing measure is closer to what a user or reviewer actually experiences, and the harness matters as much as the model.

Nvidia's board accepted the resignation of Jen-Hsun "Jensen" Huang as president, chief executive officer and member of the board on 11 November, effective immediately, according to a Form 8-K the company filed with the US Securities and Exchange Commission. Nvidia announced the change in a press release on 12 November. Neither document gives a reason or describes any continuing role for Huang.

The board appointed Colette M. Kress, Nvidia's executive vice president and chief financial officer since 2013, as interim CEO while it searches for a permanent successor, and says it will consider internal and external candidates. Kress remains CFO in the meantime. The same day, the board appointed Harvey C. Jones, a director since 1993 and a former CEO and executive chairman of Synopsys, as chair, which the release describes as the company's first-ever board chair.

News reports describe the background. On 29 October The New York Times reported that Huang had a close relationship with Alice Huang, a former executive at Megaspeed, a Singapore-based Nvidia customer, and that US intelligence saw him as the target of a Chinese spying effort; the two are not related. In a later report, on 6 November, the Times said a US intelligence memo had concluded that she was recruited by China's Ministry of State Security to gather information on him, according to The Verge. These are allegations: China's Foreign Ministry has called them "unsubstantiated", according to AP. On Sunday 8 November, Nvidia's board opened an independent investigation, overseen by a committee of independent directors assisted by outside counsel, into reports that Huang "has been the target of a foreign intelligence effort", saying it had "no reason to believe" national security or Nvidia secrets were compromised. The same day, AP reports, Huang apologised in an email to employees for being "unfaithful to my wife", without naming anyone.

On 11 November the Times, citing five people who worked at Nvidia, reported that Huang had also had an affair with an executive who reported directly to him; AP said it could not independently confirm that report. According to The Verge, The Wall Street Journal then reported that the board would meet that evening, and the board accepted Huang's resignation after a late-night meeting. AP reports that President Trump has said the FBI is looking into the relationship with Alice Huang, and The Verge notes open questions about whether it influenced Huang's reported push to sell advanced chips to China.

In the release, Jones says the board expects "the company's operations, customer commitments and product roadmap to continue without interruption." Nvidia says it will report third-quarter fiscal 2027 results on 17 November as previously scheduled.

Why it matters. Much of today's AI training and inference runs on Nvidia GPUs and its CUDA software, and Huang had led the company since he co-founded it in 1993. His exit, tied in reporting to a national-security allegation that Beijing denies, puts governance at the dominant AI chip supplier under scrutiny; AP reports lawmakers from both parties calling for further investigations. So far the only official word on continuity is the board's expectation that the product roadmap continues without interruption.

Anthropic launched the roadshow for its proposed initial public offering on 10 November. It is offering 825 million Class A shares, 801,351,354 of them from Anthropic and 23,648,646 from existing stockholders, at an expected price of $70 to $78 per share, and has applied to list on the Nasdaq Global Select Market under the symbol ANT. The underwriters would get a 30-day option to buy up to 123,750,000 more shares, and Goldman Sachs and Morgan Stanley are named as sources for the preliminary prospectus.

On those figures, the shares Anthropic itself is selling would raise between about $56 billion and $62.5 billion before fees and the underwriters' option. The company receives nothing from the shares sold by existing stockholders. Anthropic's IPO FAQ says that, subject to regulatory clearance, it expects to set the final price on 18 November and for shares to begin trading on 19 November (US Eastern time). Separately, CNBC reports, citing the company's filing, that AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC have agreed to buy $26 billion of Anthropic's Class N common stock in a private placement at the IPO price. The offering will be made only by prospectus.

Why it matters. A listing would put a frontier lab's finances, risks and compute commitments into regular public filings, giving developers and enterprise buyers a clearer view of the economics behind model prices and roadmaps.

LangChain raised $300 million in a Series C at a $3 billion valuation, announced on 13 November. Existing investor CapitalG led, with new investors GIC and DFJ Growth and existing investors IVP and Sapphire Ventures. Salesforce Ventures, Insight Venture Partners, QuantumLight and MongoDB Ventures also joined, and Sequoia, Amplify and Benchmark invested again.

The company says more than 7,000 organisations use LangSmith, its framework-agnostic platform for running, improving and governing agents. It groups the product into three parts: Agent Runtime (Deep Agents, LangGraph, LangChain, Agent Server, Sandboxes, Managed Agents and Fleet), Agent Improvement (observability, evaluation, the LangSmith Engine agent and fine-tuning from traces) and Agent Governance (gateways for models and tools). LangChain says it will announce new LangSmith capabilities on 20 November. Customer figures, such as Workday cutting the time to deploy an agent from five months to under a day, are the company's own.

Why it matters. The money is going to the layer around the model: runtime, evaluation and governance. That is where enterprises struggle to move agents from demo to production, and where platform vendors, including the model labs, now compete.

Fractile, an AI chip startup based in London and Bristol, announced a Series C on 11 November. The UK government, whose Sovereign AI Venture Fund invested alongside the British Business Bank, describes the round as worth more than £500 million at a valuation of over £5 billion. TechCrunch reports the round at $720 million, led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint, and cites Financial Times and Reuters reports of a valuation around $7 billion. The British government lists CapitalG among the other investors.

Fractile's design stacks DRAM directly on top of its chips, aiming to get around the usual trade-off between memory capacity and speed in inference, and it is building a machine-learning compiler alongside the hardware. In a LinkedIn post quoted by TechCrunch, chief executive Walter Goodwin said the funding will let the company ship its first chips and systems next year. No performance figures for the hardware have been published.

Why it matters. Memory bandwidth and inference cost, more than training, now limit long-running agents. Purpose-built inference chips are one of the main bets on lowering that cost, and governments increasingly treat them as strategic industry.

OpenAI introduced ChatGPT Wallet on 13 November, a payments feature that lets its dots agents complete a purchase after the user reviews and approves it. OpenAI says it built the product with American Express, Mastercard and Visa. After approval, the payment network issues a payment token — a substitute card number, security code and expiration date — so the user's actual card details are not shared with the agent or the merchant, according to OpenAI. Users keep their card's existing rewards and protections for eligible disputes.

Availability is narrow. OpenAI's help centre says Wallet is rolling out gradually through dots, and even eligible users may not have access yet. It is initially limited to ChatGPT Pro subscribers aged 18 or older in the United States, with a US billing address and an eligible US-issued American Express, Chase, Mastercard or Visa credit card. Debit cards and virtual cards are not supported at launch. It works only in a conversation with a dot, not in Chat, Work or standalone Codex threads, and each approval requires biometric verification such as Face ID or Touch ID on a supported device. OpenAI also notes that some websites may block agents, and that approving a purchase does not confirm the merchant has placed the order. The Verge and CNBC report the same launch date, partners and US Pro limits.

Why it matters. This is a widely visible attempt to let a general-purpose agent spend money, with a human checkpoint and tokenised credentials as the safeguard. The useful design question is the permission boundary: who approved what, for which merchant and amount.