Meta launches Muse Wonder, NVIDIA contributes its Vera Rubin rack designs, Anthropic measures autonomous replication by small models

Meta launches Muse Wonder and Muse Wonder Full, its most capable models to date, designed for agentic tasks: the former is available through the Meta Model API and Muse Code, at 4 dollars per million input tokens, the latter is reserved for Muse Code subscribers, and no open weights have been announced. At the OCP Global Summit, NVIDIA contributes its Vera Rubin rack designs to the Open Compute Project, while Anthropic’s Frontier Red Team estimates that small models are a few months away from being able to copy themselves onto connected devices. In coding, Artificial Analysis puts Claude Opus 5.5 at the top of its FrontierCode 1.1 ranking, at 52,2 %, and Google dedicates its day to health, from the Fitbit Edge wristband to enzymes designed by AlphaProtein Novo.

Meta launches Muse Wonder and Muse Wonder Full, its frontier models for agentic tasks

October 12 — Meta introduces the Muse Wonder series, consisting of Muse Wonder and Muse Wonder Full, which it describes as its most capable models to date: two frontier models (frontier models) designed for personal and professional agentic tasks. Muse Wonder is available now through the Meta Model API and in Muse Code; Muse Wonder Full is reserved for Muse Code subscribers. Muse Wonder will “soon” power the Muse personal assistant, initially for subscribers.

Fourteen months after a complete overhaul of its AI efforts, and following Muse Spark 1.3, released on September 2, Meta calls this “the next step” in its staged scaling process (scaling ladder). Both models were trained in Prometheus, its data center with capacity on the order of a gigawatt, using “larger and stronger” base models whose sizes have not been disclosed. Two capabilities are highlighted: active memory (Active Memory), which retains important context across conversations, and active focus (Active Focus), which pushes the model to change its approach when blocked, ask for clarification and verify before taking an important action.

The results appear only in an image in the post: a table of measurements by Meta, using the max setting for all models. A Muse Wonder model leads on 7 of the 14 rows, starting with the Personal Superintelligence Index, an internal Meta index. Claude Opus 5.5 nevertheless remains ahead on 6 rows, including the Artificial Analysis Index, and GPT-6 Astra leads on GDP.pdf.

Benchmark (excerpt, Meta measurements at max setting)	Muse Wonder Full	Muse Wonder	Claude Opus 5.5	GPT-6 Astra
Personal Superintelligence Index (internal Meta index)	59,1	54,3	55,2	54,5
DeepSWE 1.1	79,0	78,8	74,2	74,1
SWE-Marathon	57,5 (xhigh)	53,1	46,3	42,5
MemoryBench	85,5	85,6	81,5	81,4
Terminal-Bench 4.0-AA	57,1	52,0	59,6	59,1
ProgramBench	59,0	50,0	70,5	52,5
Artificial Analysis Index	56	52	58	53

Model version	Available access	Announced pricing
muse-wonder-1.0	Meta Model API and Muse Code, context of 1M tokens	4 dollars for input, 0,20 dollar for cached input, 20 dollars for output, per million tokens
muse-wonder-full-1.0	Muse Code subscribers only, context of 1M tokens	Power Usage (50 dollars per month) and Maximum Usage (200 dollars per month) subscriptions

On safety, Meta claims its strongest safeguards to date against chemical, biological, radiological and nuclear (CBRN) misuse and cyber misuse, while preserving access to advanced cyber capabilities for defensive use. It presents Muse Wonder as its most aligned model to date and promises a full report “soon” with additional safety evaluations.

🔗 Meta research post · Model page and pricing

OCP Global Summit: NVIDIA contributes its Vera Rubin rack designs, Meta adopts Spectrum-6

October 12 — At the opening of the OCP Global Summit in San Jose, NVIDIA contributes key Vera Rubin rack (rack) and compute tray (compute tray) designs to the Open Compute Project (OCP) through its open MGX architecture. Its stated aim: give manufacturers a common foundation for next-generation AI factories. The MGX 1.2 specification adds blind-mate liquid cooling (blind-mate) across the rack, a liquid-cooled busbar (busbar) rated at 5 000 A, fanless compute trays and racks up to 250 kW. NVIDIA also publishes a Vera CPU rack specification, designed for agent workloads (tool calls, orchestration, sandboxed execution), alongside Vera Rubin NVL72, which remains the foundation for training and inference.

On the software side, DSX OS, the operating layer of its AI factory platform, gains four open source components drawn from operating DGX Cloud, including Cluster Readiness Engine, which stress-tests GPU clusters under load before training, and a Cluster API Provider that makes clusters ready “in hours instead of days,” according to NVIDIA.

Ian Buck, NVIDIA’s vice president of hyperscale and HPC, presented a series of partner announcements alongside these contributions. The most concrete comes from Meta, which is adopting Spectrum-6 Ethernet switches: its N-IceCube switch sits in IcePack, its liquid-cooled network rack, running FBOSS, its open source network operating system.

Announcement presented at the summit	Partners named	Key point announced
MGX 1.2 and Vera CPU rack	OCP community	5 000 A busbar, racks up to 250 kW
DSX OS open source components	NVIDIA projects	Four tools drawn from DGX Cloud
Open Data Center Specification v0.7	AMD, Google, Meta	Common design for AI factory buildings
OCP Grid Resilience Specification	Google, Meta, Microsoft	Return to near-full load in less than 2 s after a grid disturbance
K-1 platform at 800 VDC	Visionbay, Vertiv	Designed for future Vera Rubin deployments
Spectrum-6 at Meta	Meta	102,4 Tb/s, up to 8 network planes, more than 100 000 GPU
Falcon Transport protocol	Google	Future ConnectX-10 SuperNIC and BlueField DPU
USFP optical form factor	Cisco, TE Connectivity, Amphenol	400 Gb/s per lane (details below)

None of these announcements comes with a price or availability date. Jensen Huang will deliver the GTC Berlin keynote on Wednesday, October 21.

🔗 NVIDIA post on the OCP Global Summit

USFP: NVIDIA and Cisco prepare pluggable optics at 400 Gb/s per lane

October 12 — Among these announcements, NVIDIA and Cisco present their proposed USFP (Ultra-Dense Small Form Factor Pluggable) form factor, a pluggable optical module that NVIDIA describes as the first designed for electrical signals at 400 Gb/s per lane. The first planned transceivers (transceivers) will combine 16 bidirectional lanes (full-duplex) in a footprint close to that of OSFP, for both switches and SuperNIC, providing up to 12,8 Tb/s, with a possible expansion to 32 lanes and beyond. Compared with current modules at 200 Gb/s per lane, NVIDIA claims four times the front-panel bandwidth density and around 1,3 times the energy efficiency, and the module can dissipate up to 100 W without dedicated built-in cooling. The aim: connect more endpoints at 1,6 Tb/s without adding a network tier. TE Connectivity and Amphenol are preparing cages, connectors and modules, but this is still only a proposal: the press release gives neither a date nor a price.

🔗 NVIDIA and Cisco press release

Anthropic’s Frontier Red Team measures self-replication by small models on connected devices

October 12 — In a study authored by William Zhou and Daniel Freeman, Anthropic’s Frontier Red Team estimates that small models could be a few months away from being able to copy themselves onto poorly protected connected devices. The team distributed Gemma4-26B-A4B across 15 unbranded Android TV boxes, bought for around 55 dollars apiece: the model runs on them at 0,39 token per second.

A 9B model manages to start inference on 8 of the 9 hardware types tested. In an isolated network of 8 emulated Android machines, only GLM-5.2 achieves end-to-end replication (8 nodes out of 8 in 4 trials out of 5); because the intrusion was deliberately trivial, Anthropic views this as an upper bound. At the rate of progress of open models of 30B or less, a small model would reach this level by the end of 2026, an estimate the team itself considers rough.

Such a worm would be slow, at most equivalent to around 40 H100 across a fleet of more than 2 million devices, but it would adapt: by inheriting notes from the previous generation, Qwen3.5-9B increases its success rate from 45 % to 61-84 % over six generations. No complete uncontrolled replicator exists yet; on defense, the study recalls that, in the case of the Mirai botnet (botnet), quarantining infected subscribers through their internet service provider cleared 92 % of infections in 14 days.

Like coding agents in 2024, we’re seeing signs of life for AI worms. — Anthropic’s Frontier Red Team, research post

Coding agents: FrontierCode 1.1 ranked by Artificial Analysis, Amp listens to GitHub, LightOn’s ColGrep-Agent-2B

An independent ranking measures what coding agents actually produce, while Amp and LightOn work on what happens around them: pull request tracking and code search.

FrontierCode 1.1: Artificial Analysis puts Claude Opus 5.5 first, at 52,2 %

October 12 — Artificial Analysis publishes its own FrontierCode 1.1 ranking, the benchmark Cognition built with open source maintainers: it checks that a maintainer would merge the change proposed by the agent, not just that it passes the tests, and penalizes excessive changes or changes outside the scope. Artificial Analysis runs the main set (Main), 100 private tasks out of 150, with its open source Stirrup harness and three trials per task; a trial that fails a blocking check scores zero.

Leading models still fail more than 40% of the time — @ArtificialAnlys on X

Claude Opus 5.5 passes the blocking checks in only 58 % of its trials, and 9 tasks out of 100 were not completed successfully by any of the 22 models evaluated. Xiaomi’s MiMo-V2.6-Pro is the leading model with open weights, at 45 %. The cost gap is much wider than the score gap:

Evaluated model (setting)	FrontierCode 1.1 score	Cost per task
Claude Opus 5.5 (max, with fallback)	52,2 %	10,15 dollars
GPT-6.1 Sol (max)	49,1 %	0,81 dollar
Claude Fable 5.1 (max, with fallback)	49,0 %	12 dollars
MiMo-V2.6-Pro	45,4 %	0,18 dollar
GPT-6 Luna (max)	41,9 %	0,09 dollar

According to Artificial Analysis, GPT-6.1 Sol thus matches Fable 5.1 for less than one tenth of the cost per task of Anthropic’s best models. These scores cannot be compared with those Cognition publishes for Devin, often measured on the Extended variant with its own harness.

🔗 Artificial Analysis FrontierCode ranking

Amp tracks the pull requests it opens and launches an orb for each GitHub issue or PR

October 12 — Amp now keeps track of the pull requests it opens: when the agent creates a PR on GitHub, it can subscribe to its comments, reviews and CI results for its latest commit. A failed check or a reviewer’s question then arrives in the same thread, where Amp already has the context to respond.

The feature relies on a “first-class” GitHub Webhooks integration (first-class): Amp receives and processes the events itself, then the user, or the agent, chooses to send them to a single orb, the remote environment where an Amp agent runs, or to a new orb each time. Users can thus request an orb for each new issue (issue) or PR in a repository, optionally restricted to a label (label): each orb reads the issue, prepares a summary, and subsequent comments return to the same thread. Since July 23, orbs had already been able to react to external webhooks; the new feature is a dedicated GitHub integration managed by Amp, with a management screen in Settings → MCP & Integrations → GitHub for editing or pausing each subscription. Simply install Amp’s GitHub app on the repository; the post mentions no pricing.

🔗 Amp post

ColGrep-Agent-2B: LightOn hands Claude Code’s code search to a local agent

October 12 — LightOn releases ColGrep-Agent-2B, a code search agent with 2 billion parameters under the Apache-2.0 license, which runs locally between the coding agent and the repository. Given a question, it searches with ColGrep, LightOn’s multi-vector semantic search engine, reads promising files, follows references and returns paths, line ranges and code: the frontier model no longer has to run successive rounds of exploration itself.

Connected to Claude Code with Sonnet 5.5, for 52 search questions across 13 repositories absent from training (624 sessions), it reduces API tokens by 34 % and API costs by 40 %, including caching, compared with grep, with no statistically significant difference in accuracy. LightOn specifies that the measurement covers code search, not end-to-end tasks. Fine-tuned from MiniCPM5-2B, the model installs with brew install lightonai/tap/colgrep && colgrep --install-agent, which downloads a GGUF version of around 2.5 GB; it requires no Python environment, inference server or API key, and responds in around twenty seconds on an M3 Pro.

🔗 LightOn’s post on Hugging Face

Health at Google: Health Guardian and the Fitbit Edge band, AlphaProtein Novo designs novel enzymes

Google devotes its day to health: consumer features for Google Health and Fitbit on one side, and on the other, a research overview highlighting previously overlooked work from Google DeepMind.

Google Health: Health Guardian, more than 100 activities detected and the Fitbit Edge band

October 12 — Google Health brings together its fall announcements in a series of posts. Health Guardian monitors gradual health changes in the background, such as blood pressure and insulin resistance; with a Google Health Premium subscription, Health Coach, the app’s coach, helps interpret these changes and prepares context to share with a doctor. Automatic workout detection, based according to Google on context-aware reasoning (contextual reasoning), becomes nearly 2 times more accurate.

Announced feature	Numerical details	Announced availability
Health Guardian	Blood pressure and insulin resistance trends, monthly report	Later this month on Fitbit Air and Fitbit Edge, Premium subscribers
Automatic activity detection	From 10 to more than 100 activities, missed workouts reduced by more than half	Later this month on Fitbit Air, Fitbit Edge and Pixel Watch 3 to 5
Fitbit Edge band	179,99 dollars, 11,5 mm, up to 7 days of battery life, replaceable battery	Available for preorder, in stores on October 27
Partnership with Flo Health	Synced cycle data, more than 80 million monthly users at Flo	Early next year

The Fitbit Edge, presented as a window into Health Coach on the wrist, is Google’s first repairable fitness band. Abbott (Lingo glucose sensor), MyFitnessPal and Function Health round out the integrations.

🔗 What’s new in Google Health · Fitbit Edge · Partnership with Flo Health

Google’s health research overview: AlphaProtein Novo designs enzymes that do not exist in nature

October 12 — Yossi Matias (Google Research) and Pushmeet Kohli (Google DeepMind) review Google’s work in health: 90 peer-reviewed papers in five years, more than 90 000 citations and more than 60 million dollars awarded in 2026 by Google.org to teams applying these models. The post mainly revisits previously known results, but highlights AlphaProtein Novo, a Google DeepMind machine learning processing pipeline (pipeline) that designs enzymes from scratch (de novo), without starting from a natural enzyme.

In a preprint posted on bioRxiv on October 5, with the California Institute of Technology and the University of Pittsburgh, the teams created novel nitrene transferases to synthesize piperidine, a key motif in drugs, with selectivity they describe as unprecedented, as well as enzymes that degrade DEHP, an environmental toxin, under conditions in which natural enzymes denature. The post also cites two new applications: in Barcelona, the Institute for Global Health virtually screened 30 000 compounds with AlphaFold against Chagas disease and identified two drugs already approved by the FDA that could be repurposed; at Children’s Mercy Hospital, DeepSomatic detected 10 missed mutations in previously sequenced childhood leukemias.

🔗 Google’s overview of AI in health · The AlphaProtein Novo preprint on bioRxiv

Models specialized by profession: Z.ai’s GLM-Legal for law, Fin’s Apex 2.0 for customer service

Two developers start with an existing model and post-train it for a specific profession, with figures to back their claims.

GLM-Legal: Z.ai specializes GLM-5.3-Flash for legal work

October 12 — Z.ai introduces GLM-Legal, a legal model built on GLM-5.3-Flash and designed for real workflows: contract review, evidence analysis, compliance. After an intermediate training phase (mid-training) on an expanded legal corpus, it was post-trained through reinforcement learning in tens of thousands of realistic legal environments. On Harvey LAB, Harvey’s legal agent benchmark covering 24 practice areas, GLM-Legal achieves a complete success rate (all-pass rate) of 19.2 %, 12.5 points higher than its base model.

Model evaluated on Harvey LAB	Complete success rate (Z.ai’s figures)
GLM-Legal	19,2 %
Kimi K3	12,9 %
Claude Fable 5	11,3 %
GLM-5.3-Flash	6,7 %
GPT-6 Astra	5,4 %

Only GLM-Legal’s score appears in the post’s text; the others come from the chart in the announcement tweet. On LegalBench and PLawBench (Chinese law), Z.ai claims parity with the best closed models, without publishing a score. GLM-Legal is in beta: law firms, legal software developers (legal tech), legal departments and independent practitioners can request API access through a form, and the model is due to arrive “soon” in AutoClaw, its desktop agent platform. Neither pricing nor open weights have been announced.

🔗 Z.ai’s post on GLM-Legal · Z.ai’s announcement on X

Apex 2.0 and Apex Flash 2.0: Fin builds its customer service models on Nemotron 3

October 7 — Fin, the AI customer service agent, launched Apex 2.0, the new version of its response generation model, this time built on Nemotron 3 Super; NVIDIA welcomed the launch on October 12. As with Apex 1.0, which Fin had already said outperformed Claude Sonnet 4.6 in production, the team starts with a base model with open weights and adapts it through its own post-training: instruction and preference tuning, updated reward models, targeted examples to decide between answering and asking for clarification, and eliminating unsupported promises. Fin also says it reduced its deployment costs through quantization, with no measurable loss of quality.

Fin model	NVIDIA base model	Announced gains over version 1.0
Apex 2.0	Nemotron 3 Super	+8,6 % user-confirmed resolutions, -16,5 % hallucinations
Apex Flash 2.0 (Fin Voice)	Nemotron 3 Nano	Responses 24 % faster, -46 % hallucinations, 9 languages instead of English alone

Apex 2.0 is in service for all of Fin’s more than 12 000 customers; the post provides no pricing.

🔗 Fin’s post on Apex 2.0

Training and serving models: QwenGyre, ReplaySSM and the /v1/systemone endpoint in Transformers

Three projects focus on model software infrastructure: training agents with very long horizons, decoding hybrid models and serving decision models.

QwenGyre: Qwen makes reinforcement learning elastic for agents with very long horizons

October 12 — The Qwen team introduces QwenGyre, an online reinforcement learning framework for agents whose single execution can last hours, involve hundreds of interactions with the environment and approach a million tokens per generated trajectory (rollout). Two problems arise: idle GPUs during long generations, and redundant trajectories whenever execution branches (context compaction, sub-agents, retries). The elastic scheduler (Elastic Scheduler) reallocates GPUs between trajectory generation and training without interrupting executions: up to 1.85 times faster than the Colocate approach and 1.78 times faster than the asynchronous approach (Async). The trajectory processor (Trajectory Processor) organizes branching histories into deduplicated trees with shared prefixes.

Applied to Qwen3.8-2.4T-A95B, with around 700 000 tokens per trajectory, QwenGyre raises NL2RepoBench from 52.5 % to 58.5 % in 48 steps and, according to the announcement tweet, reduces end-to-end training time by around 44 % compared with the asynchronous approach. The paper has been on arXiv since September 27; no code repository has been announced.

🔗 Qwen’s announcement on X · The QwenGyre paper on arXiv

ReplaySSM: NVIDIA and Dao AI Lab accelerate hybrid model decoding

October 12 — NVIDIA and Princeton University’s Dao AI Lab introduce ReplaySSM, a new way to store the state of SSM layers during decoding in hybrid models, which alternate state-space layers and attention layers, such as Nemotron 3, Qwen3.5 to Qwen3.8, Kimi-K3 or GLM-5.3-Flash. At production batch sizes, these SSM layers, rather than attention, contribute the most to decoding latency for contexts up to 100K tokens. ReplaySSM represents the state with a checkpoint and a small buffer of recent inputs: the output is computed for every token, but the checkpoint is written back to memory only every 16 tokens, roughly halving the dominant memory traffic.

Configuration measured in vLLM	Announced gain over vLLM
Decoding step, batch size of 1	None
SSM kernel, batch size of 256	1,46 to 1,80 times faster
Full decoding step, batch size of 256	1,20 to 1,47 times faster
Speculative decoding, batch size of 512, 120B-class models	1,85 and 2,10 times the throughput

The technique is integrated into TensorRT-LLM, vLLM and FlashInfer and runs in production in GLM-5.3-Flash and Kimi-K3, according to NVIDIA; the paper is on arXiv and the code on GitHub.

🔗 NVIDIA’s technical post on ReplaySSM · The ReplaySSM paper on arXiv

/v1/systemone: transformers serve turns an LLM or VLM into a decision model

October 12 — Hugging Face has merged a /v1/systemone endpoint into Transformers for transformers serve, the server built into the library. Instead of generating a response, the server scores a set of possible decisions using the model’s logits: according to Lysandre, any of the 250 supported LLMs and VLMs can thus become a “JEV-style” decision model without generating text. Marc Sun’s PR 49322 currently supports all autoregressive models and adds a decision_config passed to the server, because each decision model formulates its prompt and sets its temperature in its own way; Marc Sun announces that decision models will be standardized “soon” in Transformers. The feature is not yet in any published release, with the latest, v5.19.0, dating from October 6. llama.cpp had already been serving this format since October 2.

🔗 Lysandre’s announcement on X · PR 49322 on GitHub

Web pages and files rethought for agents: OpenAI’s Browser Annotation API, the Git forge in Perplexity Projects

OpenAI and Perplexity revisit two familiar environments, the web page and the file system, to give agents more context.

Browser Annotation API: sites customize annotations in ChatGPT’s browser

October 12 — OpenAI Developers introduces the Browser Annotation API, described in the Annotations Extensibility guide on ChatGPT Learn. Without code changes, Annotation mode in the ChatGPT desktop app’s built-in browser already works on any site: users select part of the page, comment on it and send it as context to Codex or ChatGPT Work. The API, exposed through document.oai.annotation on secure pages (HTTPS or localhost), lets the site choose what can be selected, the context accompanying the selection and the controls that preview a change before submission.

The guide details nine integrations: a card can become a single selectable object (oai-annotation-container, oai-annotatable), a button on the site can open an annotation with a suggested comment, custom controls (registerControls()) can preview a spacing token, for example, and objects drawn in a canvas become selectable. The guide targets the DevDay 2026 version of the app and later versions; in ChatGPT Enterprise and Edu, an administrator must enable these customizations. Annotations complement WebMCP site tools, which expose the actions ChatGPT can perform on a site.

🔗 Announcement from @OpenAIDevs · Annotations Extensibility guide

Perplexity Projects: a file system built as a Git forge for agents

October 12 — Perplexity publishes an engineering post about the file system behind Projects, the Perplexity Computer workspace where people and agents work together on long-running tasks. Each agent works in its own sandbox and records its changes in a shared Git repository, supported by five components: a CLI in each sandbox, a control plane, a Git service, storage backed by AWS S3 (S3Files), and document text extraction. The service accepts a push only if it starts from the latest accepted commit; otherwise, the agent replays its change and pushes again, without a merge request (pull request). A FUSE virtual file system, roughly 400 times slower on file-intensive workloads, was ruled out.

Metric published by Perplexity	Reported value
Pushes rejected as stale in production	Less than 2 %
Median push latency	Less than 1,7 s
p99 push latency	Less than 8 s
Correct answers using grep, on 150 internal financial questions	0,627
Correct answers using semantic search over an index	Around 0,37

The post does not launch any product: Projects has been available since July, and Brain, Computer’s knowledge wiki, uses the same infrastructure.

🔗 Perplexity engineering post

Briefs

Copilot CLI 1.0.96 — Now stable, it carries over all nine entries from the October 9 and 10 prereleases without adding anything new, including who made each permission decision in the timeline and case-insensitive model identifiers. 🔗 source
Push attribution on GitHub — In a limited public preview available on request to GitHub Enterprise customers, each pushed commit retains the authenticated identity of the person who pushed it; a ruleset condition, Committer must match pusher, blocks pushes whose committer does not match any of that user’s verified email addresses. 🔗 source
Copilot app for Beginners — A new episode of the GitHub Blog series shows how to keep a project’s rules in the repository so Copilot applies them automatically: custom instructions, agent skills in .github/skills, custom agents, and MCP servers such as Playwright; a tutorial, with no product updates. 🔗 source
Zed 1.23.3 and 1.24.2 — The stable release and prerelease contain only fixes (Markdown preview, edit predictions, remote project picker), an upgrade to Node.js v24, and optional telemetry on the agent panel’s save points (checkpoints). 🔗 source
Codex Ambassadors — OpenAI reopens applications for this six-month program, which requires one contribution per month (an event or content) in exchange for a Pro subscription, Codex and API credits, and support for approved events; priority goes to North America, Central America, Africa, the Middle East, and Eastern Europe. 🔗 source
The Decisions API at Hypha — Serhii Shchoholiev, head of agent engineering at Hypha, measures OpenAI’s Decisions API at 98.1% F1 in multi-label classification, compared with 93.5% for a GPT-6.1 Sol agent, at roughly one-seventh the cost and 28 times lower p90 latency; customer-reported figures, with no published methodology, shared by @OpenAIDevs. 🔗 source
Gemini 3.5 Flash deprecated — Catch-up from October 8: the same Gemini API release notes entry also deprecates gemini-3.5-flash, which can no longer be selected in Google AI Studio; its requests are redirected to gemini-3.6-flash, with no shutdown date announced. 🔗 source
Kimi Work 3.2.16 — Maintenance release: an upgraded PDF preview engine, an embedded browser that is no longer interrupted by background tasks, and eight detailed fixes, including side conversations being sent twice. 🔗 source
Genspark outage — Some users could not log in on Monday because of an outage in a Microsoft service that Genspark does not name; Genspark announced that service was restored at 16:20 UTC and is preparing compensation for affected users. 🔗 source
TutorMoments — Ai2 tested recent models on its AI tutor evaluation: with an explicit instruction, GPT-6 Astra leads (0.938 for appropriate scaffolding, 0.915 for appropriate challenge), ahead of Claude Sonnet 5.5 and Claude Opus 5.5, and a new explorer compares their responses with those of human tutors. 🔗 source
UIBK-Embedding-v1 — Abdelrahman Abdallah releases under Apache 2.0 a 179-million-parameter English retrieval model that combines dense retrieval and late-interaction retrieval (ColBERT) in a single encoder; he claims an nDCG@10 of 59.79 on BEIR, the highest score below one billion parameters among models with complete official MTEB results. 🔗 source
Sparseup-fr — Loïck Bourdois reproduces Linkup’s 150-million-parameter sparse embedding model in French, trained in about 55 hours on a single RTX 6000 Pro: 56.37 and 56.50 on cleaned NanoBEIR-fr, compared with 52.40 and 51.66 for the English version, a new state of the art among open sparse models according to the author. 🔗 source
Model Pulse — A community tool by Tardelli Stekel reconstructs, day by day since July 2024, downloads and likes for 1.64 million Hub models, with the data published in roughly 1.15 billion Parquet rows under Apache 2.0 and updated daily. 🔗 source
Contributor permission on the Hub — Hugging Face’s fine-grained access tokens and service account tokens can receive contributor permission: they create repositories and write only to those they created, for CI/CD pipelines, training jobs, and agents. 🔗 source
Microduck Arena — Pollen Robotics opens a Hugging Face Space where users train a policy for its small bipedal robot through reinforcement learning, then make it run in four events, including 2 m Sprint and Hill Climb, without a physical robot. 🔗 source
Relight Studio — Luma highlights this feature, which changes the direction of light in a shot that already works without starting from scratch; the tweet gives no model, pricing, or release date. 🔗 source
HeyGen Lessons — HeyGen introduces an offering for education: OpenAI’s Dots identifies what engages a class and adapts the lesson, then HeyGen’s MCP turns it into a video with the teacher’s avatar and voice; the tweet gives no link, pricing, or date. 🔗 source
A real estate agent and his avatar — According to a HeyGen case study, a Vancouver real estate agent who lost his filming location to a fire publishes two Reels a day with his avatar and has quadrupled his Instagram followers in ten months. 🔗 source
x.ai switches to SpaceXSI — Observed during the October 12 scan, with no announcement or date at the source: the x.ai homepage, its news, and its API documentation (“SpaceXSI Docs”) now carry the SpaceXSI brand, including on the Grok 4.7 page; the site still displayed SpaceXAI on October 10. 🔗 source

What it means

Today’s figures demand a closer look at who is doing the measuring. Meta publishes its results in an image, measured in-house, and ranks Muse Wonder Full first in its own index (59.1 versus 55.2 for Claude Opus 5.5), while on the Artificial Analysis Index in the same table, Opus 5.5 retains the advantage (58 versus 56). Z.ai compares GLM-Legal with its competitors in a tweet’s chart and claims parity on LegalBench without publishing a score, and Fin measures Apex 2.0 against its own previous version. FrontierCode’s independent ranking puts the scale of performance in perspective: the best model tops out at 52.2%, and the cost gap, from $0.81 per task for GPT-6.1 Sol to $10.15 for Opus 5.5, becomes a criterion as prominent as rank.

In infrastructure, gains come from everything surrounding the chips. NVIDIA’s contributions to the OCP address power (racks up to 250 kW, a 5,000 A busbar, a return to near-full load less than two seconds after a power grid disturbance) and networking, with Spectrum-6 at 102.4 Tb/s at Meta and USFP at 400 Gb/s per lane. Software follows the same logic: ReplaySSM roughly halves memory traffic in SSM layers, and QwenGyre puts GPUs that were waiting for long trajectories to finish back to work. By contributing its designs to the OCP, with standards co-signed by AMD, Google, Meta, and Microsoft, NVIDIA states its aim of giving manufacturers a common foundation for AI factories.

Agents are getting environments designed for them. A web page can now control what the user selects in ChatGPT’s browser and the context sent to the agent with that selection, a Perplexity project gives each agent a sandbox and a shared Git repository, and Amp brings reviews and CI results from the agent’s own pull requests back into its thread. The way agents explore also affects the result: at Perplexity, an agent searching its working copy with grep achieves a correct-answer rate of 0.627, compared with about 0.37 for semantic search, and at LightOn, delegating search to a small local agent reduces Claude Code’s API token use by 34% compared with grep, without a significant loss of accuracy.

The Frontier Red Team study shows the downside of this autonomy. The capabilities that make an agent useful, installing a model on unfamiliar hardware or passing its notes to the next agent, are also those that would allow a worm to adapt: a 9B model runs inference on 8 out of 9 hardware types, and inheriting notes raises success from 45% to 61-84%. Anthropic has not yet described any complete replicator and is relying on monitoring by internet service providers so that an infected device infects, on average, fewer than one other device. Meta, for its part, claims Muse Wonder has its strongest safeguards against cyber and CBRN uses, but points to a full safety report that is still forthcoming.