Vibeleaderboard
Index — Latest Intelligence

Intel

Page 16
Analysis & explainers →

Arete 27B released

Nous Research releases Arete-27B, an SFT-only post-train of Qwen3.8-27B under Apache 2.0.

Why it mattersAn open 27B model fine-tuned to shift value framing, with self-reported evaluation. It shows how SFT alone can change a model's political framing, and gives builders an alternative base.

articleNous Research model releases

Meta Brings Muse Wonder Memory and Focus to Muse and Meta AI

Meta is rolling out specialized models trained with Muse Wonder to Muse and Meta AI, adding Active Memory and Active Focus.

Why it mattersMeta ships models trained for agent memory and recovery from blocked tasks, with self-reported gains (2.4x less forgetting of critical context, 67% better recovery). It shows how labs target long-running agent reliability.

articleAIatMeta

Accelerating Reinforcement Learning Weight Refit with the Reshard Collective

NVIDIA explains the NCCL Reshard collective, which reformulates RL weight refit as a DTensor-to-DTensor transfer.

Why it mattersRefit for the largest open models can move about 10 TB per step. Reshard transfers only the theoretical minimum data between layouts, giving 2.7x to 5.7x faster refit in NeMo RL, which cuts RL post-training overhead.

articleTanya Lenz
Artificial Analysis@ArtificialAnlys

Artificial Analysis launches VIALS-AA, a visual-reasoning benchmark for life sciences, run through its Stirrup agent harness. Top models land within 3 points (64%, 63%, 61%) at very different cost per task, and pass^3 consistency separates models more than single-attempt scores.

Why it mattersTop models sit within 3 points on VIALS-AA, but cost per task varies about 29x, driven by turns and tool calls.

Project Eve Claude Plays Mmorpgs

Anthropic put Claude in EVE Online and EVE Frontier to test agent trading and teamwork.

Why it mattersShows how Claude agents behave in long-horizon markets and teams, including profit gains across model versions and how teammates respond to a secretly sabotaging agent. Useful for designing multi-agent safeguards.

articleAnthropic research

claude-code v2.1.319

Claude Code 2.1.319 changes auto mode behavior behind LLM gateways, adds MCP task backgrounding, a denyInsteadOfPrompting permission setting, a flag-fetching kill switch, a latency-optimize command for Claude API code.

Why it mattersAuto mode now asks or denies calls the server cannot check when telemetry or flags are off, unless CLAUDE_CODE_AUTO_MODE_SERVER=0. New permissions.denyInsteadOfPrompting enables unattended runs without hanging on prompts.

repoclaude-code-publisher[bot]

claude-agent-sdk-typescript v0.3.319

Claude Agent SDK for TypeScript 0.3.319 adds background MCP task support, fixes an OpenTelemetry session count metric and session_writes_lost reporting, makes deleteSession reject on partial deletion.

Why it mattersThe Claude Agent SDK now runs MCP tasks in the background when CLAUDE_AUTO_BACKGROUND_TASKS=1 with streaming input, so long-running tool calls no longer block the session. Not available on Bedrock, Vertex or Foundry.

repoclaude-code-publisher[bot]

How Salesforce uses Claude Code

Salesforce engineering leaders describe scaling Claude Code across thousands of engineers.

Why it mattersShows how a large engineering org measures coding agent ROI and scales adoption. It covers an effective output metric built with Stanford researchers, skill sharing with evals, and replaying benchmarks to test new models.

videoClaude

Cooking with Reducto’s CEO: Why PDFs Still Break AI — Adit Abraham

Reducto's CEO explains why complex tables, handwriting and long documents still break AI, when specialized parsing models beat general ones.

Why it mattersDocument-heavy agents miss rows and misread tables even with frontier models. The discussion covers routing hard pages to specialized models and catching extraction failures in production.

videoLatent Space
Runway@runwayml

Runway added GPT Image 2.6, claiming 50% faster image generation, GIF output, SVG vector export, and precise edits that preserve characters, products and brand details. Based on Runway's announcement only; no independent benchmarks.

Why it mattersGPT Image 2.6 inside Runway adds SVG vector export and GIF output plus edits that keep characters and brand details consistent, with claimed 50% faster generation. Teams producing design assets can do this in one tool.

transformers v5.22.2 — Patch release v5.22.2

Hugging Face Transformers v5.22.2 is a patch release containing a hardening of Jinja sandboxing and a distributed fix that sets device_id only for NCCL.

Why it mattersTransformers 5.22.2 hardens its Jinja sandbox, which matters because chat templates execute from model repos. It also restricts device_id to NCCL for distributed runs. Upgrade if you load third-party templates.

repovasqu

Unity Sdk Release

ElevenLabs released an ElevenAgents SDK for Unity.

Why it mattersUnity developers can now connect a scene to an ElevenLabs voice agent through one Conversation API. It handles microphone input, spatial audio output and per-character prompt variables across Editor, desktop and WebGL builds.

R/place for Git – Open-source Git server sustaining 567 pushes/s

Reason Machines open-sourced Aspen, a git server designed for many agents pushing at once.

Why it mattersParallel coding agents can overwhelm ordinary git servers. This open-source server claims 567 pushes/s on a single 8-core machine with stock clients, relevant to anyone scaling agent CI and code storage. Figures come from the authors.

articlesvemyh

unsloth v0.1.916-beta — Large Performance Fixes

Unsloth v0.1.916-beta ships large UI performance fixes, separate and floating browsers per chat for model use, fixed unified memory offloading, up to 2x faster two-GPU image and video generation.

Why it mattersAdds per-chat browsers a model can drive, working unified-memory offload for big models, and up to 2x faster two-GPU image and video generation. Relevant if you run models locally.

reposhimmyshimmer

rf-detr 1.12.0 — v1.12.0: One prediction API and INT8 exports

RF-DETR 1.12.0 adds one RFDETRInference API for images, video and streams across native and exported models, INT8 quantized exports, faster CUDA and Apple silicon defaults, per-class thresholds.

Why it mattersRF-DETR 1.12.0 unifies inference for native and exported models, adds INT8 exports for TensorRT, ONNX and OpenVINO, and flags eight behavior changes to review before upgrading.

repoBorda

Inkling 2 Small: An Inkling of Profanity

Vals AI reads reasoning traces from Inkling 2 Small across 2,655 benchmark runs.

Why it mattersReading agent reasoning traces at scale reveals behavior signals: swearing occurs only in long runs, yet 96% of failed runs never swore. It cautions against using such markers as failure predictors.

articlewww.vals.ai

[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark

Latent Space's AINews digest on Thinking Machines' Inkling 2 Small.

Why it mattersA new Western open-weight model, Inkling 2 Small, reportedly competes with similar-size Chinese models and was pretrained quickly in native NVFP4. Engineers choosing open-weight models or fine-tuning via Tinker get a new option to evaluate.

articleswyx .ai

Accelerating Scientific Research

Anthropic profiles labs using Claude as a research collaborator.

Why it mattersShows how research labs wire Claude into tool-rich agents for genome-scale analysis and gene cluster interpretation. Useful patterns for anyone building domain agents over fragmented tools and datasets.

articleAnthropic research

7 Things Your AI Coworker Needs - Philipp Schmid, Google DeepMind

Google DeepMind's Philipp Schmid builds a finance coworker on Gemini Managed Agents.

Why it mattersShows how to give a hosted agent its own sandbox, tools and schedule while keeping credentials out of the model's reach through proxy-side substitution.

videoAI Engineer

Your AI Agent Keeps Making the Same Mistakes - Sparsh Jain, CentralAgent AI

CentralAgent AI's Sparsh Jain describes rebuilding agent memory as an experiential system.

Why it mattersOffers a memory design with note lifecycle, deactivation and scheduled forgetting to stop agents repeating mistakes, rather than just appending context.

videoAI Engineer

When Can Self-Play Fix a False Belief? Answer Keys, Pass Bands and Replacement Rules in an Execution-Grounded Loop

Expert-iteration experiments on Qwen3-4B-Base show that answer keys from the generator's own code preserve planted spec misreadings.

Why it mattersExecution-grounded self-play does not fix a false belief if the same model writes the code that produces the answer key. An independent implementation from a model without the error did correct it.

articleNirdesh Chauhan, Dushyant Rajput, Siddharth Kosaraju

js-genai v3.0.0

Google's js-genai SDK reaches 3.0.0.

Why it mattersUpgrading to 3.0.0 defaults the Gemini Developer API to v1 and requires Node.js 22 or newer. Set GOOGLE_GENAI_API_VERSION or apiVersion to keep the previous behavior.

reporelease-please[bot]

python-genai v3.0.0

Google's python-genai SDK reaches 3.0.0.

Why it mattersUpgrading to 3.0.0 changes the default Gemini Developer API version to v1 and swaps the HTTP client from httpx to httpx2. Set GOOGLE_GENAI_API_VERSION or api_version to keep the old behavior.

reporelease-please[bot]

codex rust-v0.167.0 — 0.167.0

OpenAI's Codex CLI 0.167.0 adds an agent command center, session picker backend switching, local image previews, Google Vertex AI setup, system proxy opt-in.

Why it mattersCodex 0.167.0 adds a `codex agents` command center with unread-result markers, backend switching that keeps drafts, and a Vertex AI setup wizard behind a flag.

repogithub-actions[bot]

Inkling 2 Small

Thinking Machines releases Inkling 2 Small, a multimodal mixture-of-experts model scoring 31 on the Artificial Analysis index, available on Tinker now with open Apache 2.0 weights planned after external safety testing.

Why it mattersA new 594B-parameter MoE (17B active) with image input and adjustable effort is available for sampling and fine-tuning on Tinker today. Apache 2.0 weights are promised for next week after external red-teaming.

articleThinking Machines editorial sitemap

AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue

Coatue's Frank Long frames investing as a function over outside and inside context, splits it into gather, analyze and decide steps.

Why it mattersShows how to split a knowledge-work function into steps and assign each to code, an LLM or a human, then tune against a defined objective instead of automating for its own sake.

videoAI Engineer

User Groups

Factory adds User Groups to its Organization Model.

Why it mattersEnterprise admins running Factory can now set usage limits and model access per IdP group across organizations.

articleFactory News

How we investigated prompt cache misses with Datadog

Datadog engineers trace LLM spend down to individual requests and find apps that enabled prompt caching but never benefited, because small changes between calls broke reuse or cache settings never reached the provider.

Why it mattersShows how tiny request differences and dropped cache settings silently defeat prompt caching, and how to find them by tracing per-request costs. Useful for anyone paying for repeated system prompts and tool definitions.

articlewww.datadoghq.com

Vials Aa

Artificial Analysis launches VIALS-AA, its implementation of Handshake AI's scientific image reasoning benchmark.

Why it mattersCompares frontier models on 161 scientific image tasks with three repeats each. Models with similar pass@3 differ widely on all-three-correct consistency, and cost per task varies about 99x, which affects model choice for visual agents.

articleArtificial Analysis articles

Ampersand from Humans& now available on AI Gateway

Humans& Ampersand, a model built for group conversations that decides each turn whether to speak or stay quiet, is available on AI Gateway via the AI SDK or Chat Completions, using caller-defined send_message and pass tools.

Why it mattersHumans& Ampersand is a model for multi-party conversations that chooses each turn between send_message and pass, so you can build group-chat agents that stay quiet when appropriate. Free on AI Gateway through November 13, 2026.

articleGregor Martynus
Artificial Analysis@ArtificialAnlys

Thinking Machines launched Inkling 2 Small, scoring 31 on Intelligence Index v5. It leaps on agentic tasks (FrontierCode 2.0 at 25%) but trails on Terminal-Bench and uses ~150K tokens per task. A 594B total, 17B active MoE, launched proprietary with Apache 2.0 weights planned.

Why it mattersInkling 2 Small posts big agentic gains for $1.18 per task but uses about 4x more output tokens than its predecessor. Weights are promised, not yet released, so plan around API access first.

Artificial Analysis@ArtificialAnlys

Artificial Analysis ranks Black Forest Labs' FLUX 3 Image Klein #2 among open-weight text-to-image models at $9 per 1k images. It improves on its FLUX.2 predecessors in every category, with strong layout and UI/UX results. License is free non-commercial, and commercial use is allowed under $5M annua

Why it mattersFLUX 3 Image Klein is an open-weight 5.6B image model at $0.009 per image with a free non-commercial license and commercial use under $5M revenue. It sits on the price-quality frontier for text-to-image and editing.

Nous Research@NousResearch

Nous Research ships Hermes Agent v0.23.0, the Hearth Release. Highlights include a desktop in a browser tab, one gateway for all conversations, /rewind, group chats with automatic routing, and spend limits for paid tools. Only the highlights were read, not the full notes.

Why it mattersHermes Agent v0.23.0 adds /rewind, automatic group chat routing, and spend limits for paid tools. It addresses cost control and recovery in long-running agents. Details beyond the highlights were not available.

Ricursive Ships Ripple-65B, an Open-Weight Agentic Coding Model

Ricursive released Ripple-65B Preview, an Apache 2.0 dense 65B model trained on 12T tokens and post-trained for agentic coding.

Why it mattersA permissively licensed dense 65B model tuned for terminal-based agentic coding is available at about $0.10 to $0.20 per 1M input tokens. You can test it as a cheap open-weight backend for coding agents.

articleRicursive Intelligence

OpenHands v1.36.0

OpenHands 1.36.0 focuses on automations.

Why it mattersOpenHands 1.36.0 extends its automations with template suggestions, Bitbucket Data Center and Azure DevOps repository listing, and self-hosted GitHub setup guidance. Teams running it on those platforms get easier setup.

repoopenhands-release-bot[bot]

Stop Giving Your AI Agents API Keys - Kim Maida, Keycard & Mason Egger, Temporal

Keycard and Temporal build an expense-approval agent with OAuth token exchange for per-call, user-delegated credentials and a full audit trail, then use durable execution to kill and resume the agent mid-task without re-running LLM calls.

Why it mattersShows how to scope per-call tokens to agent and user identity and make agents crash-recoverable without leaking credentials into workflow history.

videoAI Engineer

Inside GitHub’s new Git storage architecture

GitHub details its new Git storage architecture for agentic development. It minimizes coordination, orders only reference updates, and decouples storage from compute.

Why it mattersAgent fleets pushing to one repository strain Git's write path.

blogKarthik Raman
humans&@humansand

humans& says its Ampersand model's multi-agent mode, available only through its Agent API, reaches 92.2% on SWE-bench Pro v2 HARD-51 and 77.9% on DeepSWE v1.1 at lower cost. All figures are vendor-reported, from the announcement thread only.

Why it mattersA multi-agent coding mode is available through an API, with vendor-claimed frontier-level SWE-bench Pro results at lower cost. You can evaluate it against your current coding agent, treating the numbers as unverified.

How Netflix Evaluates AI When There's No Right Answer - Cameron Wolfe, Netflix

Netflix's Cameron Wolfe presents a three-layer eval framework for open-ended tasks.

Why it mattersGives a concrete way to validate LLM judges against human golden data and handle noise with power analysis, then confirm offline evals in production A/B tests.

videoAI Engineer

The Next Era Of Notion

Notion's CEO frames its fourth chapter around agents.

Why it mattersNotion reports roughly half its pages are now created by agents and is repositioning as a context layer that third-party agents read and write. Engineers building agents on workspace data should expect this to shape integration surfaces.

articleNotion Blog

Atlas Realtime

World Labs releases Frameboy, a research preview that streams Atlas world-model frames to the browser at up to 15 fps with live camera control.

Why it mattersAtlas now streams frames at up to 15 fps from cloud GPUs as you move the camera, removing the wait-per-generation loop. This makes interactive scene scouting and simulation workflows feasible to prototype.

articleWorld Labs editorial sitemap
Ant Ling@AntLingAGI

Ant Ling releases Ling-3.2-flash, a native multimodal model with 560B total and 25B active parameters. It targets visual understanding, multimodal agents and 3D tasks, and reports self-published wins over Opus 5 on OmniDocBench, ZEROBench and Vision2Web.

Why it mattersLing-3.2-flash is a 560B (25B active) multimodal model aimed at long-horizon visual agents and frontend work. Its benchmark claims versus Opus 5 are self-reported, so verify them before relying on it.

The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference

SemiAnalysis reports TPU Ironwood results on InferenceX AgentX, tracing Google's external TPU sales and upstream work with vLLM, SGLang, llm-d and PyTorch.

Why it mattersThird-party AgentX results for Google TPU show how long-context, multi-turn agentic serving compares to NVIDIA hardware on cost, which affects where inference capacity and pricing may head.

articleBryan Shan

Inside Intelligence Index v5: Evals, Progress Index and What Comes Next

Artificial Analysis walks through Intelligence Index v5 (nine evals including FrontierCode 2.0 and Terminal-Bench Science), the new Progress Index, cost and time per task.

Why it mattersIt explains what the nine Index v5 evals measure and why the Index and the new Progress Index serve different jobs. It also shows task cost varying over 1,000x and where cheap classifier models fall short.

Artificial Analysis Presents Intelligence Index v5 and Progress Index

Artificial Analysis walks through Intelligence Index v5.

Why it mattersArtificial Analysis details how Intelligence Index v5 swaps in agentic evals (Frontier Code, Terminal Bench Science) and reports cost per task spanning over 1,000x.

Is Android ready for AI agents? What broke in 3324 actions on 45 phones

Opposable analyzes 3,324 agent actions across 45 Android phones.

Why it mattersShows that mobile agents mostly fail on connectivity and input focus, not screen understanding: 108 of 305 failures were dropped phone connections. Useful when designing reliable Android agent control.

GPT-Live and GPT-Realtime: Arch Trade-Offs Behind Continuous Voice AI

An analysis of continuous voice AI trade-offs.

Why it mattersReports that a full-duplex voice frontend failed to delegate 14 of 75 cancellation requests despite explicit instructions, and argues for measuring time to first substantive information instead of time to first audio.

articleflashnik
Artificial Analysis@ArtificialAnlys

Nanbeige4.3-33B, an Apache 2.0 dense 33B reasoning model, ties Qwen3.8 27B at 20 on Intelligence Index v5. It is stronger on terminal tasks (30% Terminal-Bench 4.0), weaker on knowledge, abstains on 71% of Omniscience questions, and uses 44% more tokens per task.

Why it mattersA 33B dense Apache 2.0 model small enough to run locally scores 30% on Terminal-Bench 4.0, well above its Qwen peer. The tradeoffs are weaker knowledge and about 44% more output tokens per task.

Google AI@GoogleAI

Google released Gemini Omni 1.2 Flash, extending video generation and editing to 15 seconds, improving scene consistency and character speech, and cutting API pricing 30% to $0.07 per second at 720p. It rolls out to Flow, Vids, the Gemini app, and AI Studio.

Why it mattersVideo generation through the API now costs $0.07 per second at 720p, 30% less than the prior generation, with clips up to 15 seconds. That lowers cost and widens what can be built in one generation.

Open Model Safety Lessons

Reflection describes safety training for its open-weight Beam model and publishes Iris Safety Evaluations and an Open Safety Framework covering harm categories, prompt injection and capability-based risk of open weights.

Why it mattersReleases Iris Safety Evaluations and an Open Safety Framework that assess open-weight models for post-release safeguard removal and capability risk, giving teams reusable evals. Results details were not fully visible in the source.

articleReflection AI editorial sitemap

An index of the vibe-coding frontier. Corrections welcome.