Living records of models, tools and protocols. See the current judgment and how it got here.
vLLM 0.33 is stable and expands EFA, low-precision paths and diagnostics, but CUDA 13, ROCm and executor changes make it a non-trivial upgrade
Latest change · 2026-11-03：vLLM 0.33 moves to CUDA 13 and adds EFA serving and new inference paths
Run:ai now combines memory-enforced GPU fractioning with Workload Map reuse across Kubernetes AI workload types
Latest change · 2026-11-03：NVIDIA documents Workload Map architecture and three reuse paths
The national program is announced, but only the Georgia pilot has an approved contract and the campus is not operating
Latest change · 2026-11-03：OpenAI launches Grid Protectors with an approved Georgia pilot
Revenue scale and long-term compute commitments now appear in a regulatory filing, but interim statements are unaudited and pricing plus cash flow need follow-up
Latest change · 2026-11-03：Anthropic discloses $17 billion quarterly revenue and $895 billion of compute commitments
Reproducible checks show sampled weights can be generated from name-based seeds, with no Anthropic provenance authentication
Latest change · 2026-11-03：Community audit finds the claimed Anthropic Numbat weights are seeded noise
Version 0.26.5 moves bounded repair handoff and isolated verification into stable, while Helm managed deployment remains experimental
Latest change · 2026-11-02：Version 0.26.5 adds bounded repair handoff, isolated verification and experimental Helm deployment
The local agent harness is usable, but terminal commands are unsandboxed and network policy is heuristic
Latest change · 2026-11-02：Rapid-MLX 1.0 adds a local agent harness for Apple Silicon
One author-run experiment shows a reject-and-revise loop, but complete reproduction materials are missing
Latest change · 2026-11-02：Auctopus agents revise a quantum-spin experiment after detecting drift
Version 2.1.316 further tightens enterprise policy boundaries and fixes stability issues; adopt after regression testing
Latest change · 2026-10-31：Version 2.1.316 tightens model policy, restricted uploads and hosted safety verdicts
Version 0.166 adds queue resume and network diagnosis, while forked approvals and thread restoration still need regression tests
Latest change · 2026-10-31：Version 0.166.0 adds queue resume and repairs session-state boundaries
The 1017 release now has an NVIDIA NVFP4 package; speed and price are competitive, but total-token cost and quantized quality need testing
Latest change · 2026-10-31：NVIDIA publishes a 1017 NVFP4 checkpoint for Blackwell GPUs
Beyond the low-cost API, public 252 GB weights are now downloadable; review the custom license before self-hosting
Latest change · 2026-10-31：Follow-up verification confirms public M3.1 Flash weights, correcting the earlier unavailable status
Public weights enable local decision evaluation, but publisher benchmarks and a small translated set do not establish production reliability
Latest change · 2026-10-31：FrontiersMind releases Lumma-Fev-v2 decision models
Global English access makes watermark checks usable, but coverage and independent error-rate evidence remain limited
Latest change · 2026-10-31：Google opens SynthID Detector globally in English
The public preview makes coding-agent quality and throughput testable, but publisher results remain unreproduced
Latest change · 2026-10-31：DiffuSpace releases Clover-35B-A3B-Preview weights and code
The public blueprints compose owned agent stacks, but policy isolation and operating cost still need testing
Latest change · 2026-10-31：NVIDIA opens NemoClaw blueprints for controlled agent stacks
For anyone who installed or updated from the Cursor marketplace during the stated window, uninstalling alone is insufficient because the payload drops persistence
Latest change · 2026-10-31：Cursor reports a backdoored Claude Code extension in its marketplace
For life-science AI teams, the cases show where agents are entering real workflows beyond molecule generation
Latest change · 2026-10-31：Anthropic reports early Claude results across six drug-development partners
For video-tool teams, public weights make local quality, latency and reference control testable
Latest change · 2026-10-31：Pika Labs releases Pika H3 Flex video-model weights
Limited trusted access supports life-science research, but there are no open weights or independent reproductions
Latest change · 2026-10-31：OpenAI releases GPT-6 Rosalind Research under trusted access
Public weights and code make looped multimodal execution testable, while serving support is still landing
Latest change · 2026-10-31：SenseNova opens a looped multimodal model for long-context agents
An orbital TPU produced a data-center-class AI response, but this remains a research demonstration
Latest change · 2026-10-31：Google runs the first data-center TPU response in orbit
Version 2.2 adds persistent PTYs and experimental artifacts and schedules while fixing launch regressions
Latest change · 2026-10-31：OpenCode persists terminals and adds experimental agent artifacts
The public environment expands reproducible web-agent evaluation, while results remain author-reported
Latest change · 2026-10-31：WebAnything reconstructs reproducible environments for web agents
The public corpus and benchmark support frontier-math retrieval research, but transfer value is not independently validated
Latest change · 2026-10-31：OpenFrontiers tests whether agents rediscover or retrieve mathematics
Copilot cloud sessions are GA on paid plans; cross-surface recovery and the shared Actions entitlement merit testing, while snapshot isolation and real billing still need validation
Latest change · 2026-10-30：GitHub makes connected Copilot cloud sessions generally available
Ollama 0.40.13 expands LFM3 audio and omni paths to Linux and Windows, suitable for a cross-platform trial
Latest change · 2026-10-30：Ollama 0.40.13 expands LFM3 audio and omni models to Linux and Windows
Large 4 weights are public with named serving paths; the MaaS license restriction and 1T-model cost still need review
Latest change · 2026-10-30：Mistral releases Large 4 weights under a license that restricts MaaS
Version 4.1.1 demonstrates independent adapter updates and ten-language ASR, but the Adapter SDK remains unavailable and quality gains are vendor-reported
Latest change · 2026-10-30：Granite Switch 4.1.1 uses backward-compatible adapters to extend ASR to ten languages
Currently pre-released only to selected cyber defenders, with no general API or weights; vendor benchmarks and pricing await external validation
Latest change · 2026-10-30：Google restricts Gemini 4 Carbon to selected cyber defenders
The public library moves agent action constraints into runtime protocols; its single airline-domain result is useful but not evidence of general reliability
Latest change · 2026-10-30：Concordat enforces typed action protocols for LLM agents
The open Swift packages reduce Apple on-device speech and typed-decision integration work, but platform requirements and the Core AI dependency conflict need validation
Latest change · 2026-10-30：AnySpeechRecognizer and AnyDecisionModel open Apple-native AI components
Gemini CLI 0.66 adds private-network MCP OAuth and non-interactive safe Git, but the controls are opt-in and environment-specific
Latest change · 2026-10-28：Gemini CLI 0.66 adds private-network MCP OAuth and safer Git execution
Nemotron 3.5 Transcribe weights and Granary v2 training data are open; multilingual and far-field results are vendor-reported and need testing on target audio
Latest change · 2026-10-28：NVIDIA opens Nemotron 3.5 Transcribe and Granary v2
The research preview shows interactive world generation, but access is limited and there are no open weights or general API
Latest change · 2026-10-28：Google opens Genie 4 to trusted testers
Public code and data can test whether research agents repeat existing ideas; weak labels and domain coverage still require human review
Latest change · 2026-10-28：Novelty Bench exposes brittle LLM research judges
Controller weights and research materials are public and reproducible, but sim-to-real transfer and hardware safety need local validation
Latest change · 2026-10-28：NVIDIA opens Metatron humanoid controller weights
OpenHands 1.34 binds runtimes to conversations and extends recovery paths, making it suitable for isolated long-task lifecycle tests
Latest change · 2026-10-28：OpenHands 1.34 isolates each conversation in its own runtime
The GPT-6 family now includes lower-cost GPT-6.1 Terra and an internal Ultrafast incident-response case; performance and savings remain primarily OpenAI-reported
Latest change · 2026-10-27：OpenAI reports an internal incident-response use case for Ultrafast
Mission Control 2.5 makes Kubernetes recovery available as a commercial capability for controlled evaluation by existing large-GPU-cluster customers
Latest change · 2026-10-27：NVIDIA Mission Control 2.5 brings automated recovery to Kubernetes
For coding-agent users, the release exposes what changed and hardens several failure paths that can otherwise duplicate work or strand a turn
Latest change · 2026-10-27：Cline makes changed files and failure recovery visible across desktop and extension
For teams expressing agent behavior as GitHub Actions, secret flow across steps is a concrete safety boundary rather than a generic scanner rule
Latest change · 2026-10-27：GitHub Agentic Workflows traces secret leaks through step outputs
For teams training models to write reports, plans or documents, section-scoped constraints are closer to production than whole-answer style checks
Latest change · 2026-10-27：AI21 opens LIFT-Bench for scoped long-form instruction following
State-based checks can expose whether a multimodal agent actually changed the environment rather than narrated success
Latest change · 2026-10-27：Meta opens InteractiveWorld and a state-verifiable multimodal benchmark
For teams evaluating local document tooling, the multi-surface architecture is interesting but current community attention is ahead of release maturity and evidence
Latest change · 2026-10-27：PaperPilot draws attention before publishing a stable release or benchmarks
TypeScript SDK 2.4 adds experimental Skills and tighter validation; trial it in isolation before broad rollout
Latest change · 2026-10-25：LiteLLM fixes proxy startup failure when the cost map is offline
For teams operating multiple coding agents, permission boundaries and cost attribution now share one control plane
Latest change · 2026-10-25：Paperclip adds auditable private work and exact agent cost controls
For teams that need local long-context and coding capability, this is a directly downloadable candidate
Latest change · 2026-10-25：Nanbeige releases an open 33B model with a 512K context window
The artifacts turn multilingual training-order claims into reproducible experiments
Latest change · 2026-10-25：Beetle publishes a multilingual suite for testing learning curricula
For bioinformatics researchers, local weights can keep sensitive sequences off third-party services and lower experimentation cost
Latest change · 2026-10-25：Carbon-X opens a DNA-and-language model for local research
This may improve batching and prefill economics on older consumer GPUs, but evidence is author-reported
Latest change · 2026-10-25：GPU-Unlock exposes faster tensor paths on Ampere GeForce cards
This makes failure-ledger retries reproducible and useful for testing agent state design
Latest change · 2026-10-25：GVS5H opens a ledger-based harness for repeated coding-agent passes
The standard has moved to neutral AAIF governance with strong cross-platform adoption signals; versioning and security processes still need implementation
Latest change · 2026-10-24：Anthropic donates Agent Skills to the Agentic AI Foundation
The API is GA with clear batching, caching, pricing and data controls, making it suitable for routing and classification pilots
Latest change · 2026-10-24：OpenAI makes the Decisions API generally available
The paper shows that truthful but incomplete tool output can induce unauthorized action
Latest change · 2026-10-24：EpiSEM-Bench exposes unsafe agent actions when tools omit critical evidence
Deployment-specific simulation beats a cross-environment baseline in one environment, but public code and independent reproduction are absent
Latest change · 2026-10-24：Simulated Control Evals predicts deployment-specific safeguard failures
Public proofs are checkable, but the guarantee is narrower than a full software stack
Latest change · 2026-10-24：Anthropic and Theorem publish machine-checked proofs for C libraries
Weights and code are public, supporting a trial on real images
Latest change · 2026-10-24：Tencent releases WeVisOne-VL weights for unified visual perception
PHP has published critical security updates; affected runtimes should upgrade immediately
Latest change · 2026-10-24：PHP patches two critical core-engine flaws found with frontier models
Weights, code and training recipes are public, but evaluation remains publisher-run
Latest change · 2026-10-24：Tencent releases Hy-Nomas for interactive world generation
The public implementation is testable, but gains need reproduction on local models
Latest change · 2026-10-24：NVIDIA opens accuracy-aware tuning for TensorRT build routes
The small study is suggestive, but no public system is available for engineering validation
Latest change · 2026-10-24：Ai2 tests LitPivot for literature-grounded research pivots
The API is callable, but weights are pending and performance is publisher-reported
Latest change · 2026-10-24：Moonshot launches Kimi K3.1 Pro before releasing its weights
Only a research preview and selected demos are available; no testable artifact exists
Latest change · 2026-10-24：Block previews Dessa, a speech-only full-duplex conversation model
The demo is a useful illustration of extreme model compression, but it does not yet support a reusable product decision
Latest change · 2026-10-24：Simulated Control Evals predicts deployment-specific safeguard failures
Source is public, while official materials still conflict on whether full runtime access is public or closed beta
Latest change · 2026-10-23：TorchNeuron source is public while service access remains unclear
Human-directed search and formal verification produced a strong result, but reusable tooling and formal acceptance remain incomplete
Latest change · 2026-10-23：Researchers prove the KLS conjecture with human-directed AI search
Public code and MCP access make agent-authored interactive charts directly testable
Latest change · 2026-10-23：Microsoft open-sources Flint for agent-authored interactive charts
Open weights and code are testable, while vendor benchmarks and the custom license need review
Latest change · 2026-10-23：NVIDIA releases Kumo Relational for zero-training database prediction
The failure mode is directly relevant to RAG, but code and independent reproduction are absent
Latest change · 2026-10-23：Study finds document converters erase cues that policy agents need
Global embeddings and training artifacts are testable, while licensing and independent accuracy checks remain gates
Latest change · 2026-10-23：Ai2 releases OlmoEarth v1.3 and global 10-meter embeddings
The policy is not yet effective, but it changes architecture and contract boundaries for direct model-access intermediaries
Latest change · 2026-10-23：OpenAI announces prior review for services that resell direct model access
Apache-2.0 weights and serving paths are testable, but benchmarks remain publisher-run
Latest change · 2026-10-23：Recursive releases Flint as an open-weight sparse coding model
The proactive workflow merits watching, but access is a US Pro/Max waitlist preview and reliability and permission boundaries are unverified
Latest change · 2026-10-23：Anthropic opens the Claude Duet proactive agent as a US waitlist preview
The API, regions and prices are explicit enough for a real-audio trial; performance remains vendor-reported
Latest change · 2026-10-21：Mistral launches Voxtral 3 for production voice agents
Public code supports a one-day reproduction, but the small author-run evaluation does not establish live customer-service gains
Latest change · 2026-10-21：ALTK-Evolve adds learned memory to customer-service agents
Public tests add a baseline for lower-resource programming languages, but vendor results do not replace repository-level migration validation
Latest change · 2026-10-21：OpenAI publishes MultiLangBench for low-resource coding
Production mechanisms and scripts are public, but the first release only supports offline inference on narrow hardware and omits complete-serving components
Latest change · 2026-10-21：DeepSeek open-sources its production inference engine DeepInfer
The API is live with enough free allowance for a trial, while performance and latency still need same-condition reproduction
Latest change · 2026-10-21：Inception launches Mercury 3 as a diffusion language model
The integration remains request-access only; wait for GA details on regions, pricing and data boundaries
Latest change · 2026-10-21：OpenAI plans native model and Codex access through Google Cloud
Code and corpus enable workflow review, while biological leads still require experiments
Latest change · 2026-10-21：OpenAI publishes a distributed-agent workflow for biological discovery
OpenShell is testable now, while Sentry remains a hardware layer planned for next year
Latest change · 2026-10-21：NVIDIA publishes an OpenShell architecture and previews Sentry
Open weights and permissive licensing support a local speech trial; vendor WER and speed need real-recording checks
Latest change · 2026-10-21：IBM releases open Granite Speech 5.0 models
The public core split can expose unstable criteria behind final answers, within one fairness definition
Latest change · 2026-10-21：Ai2 releases FairStress to test hidden asymmetry in LLM decisions
The organization-level npm proxy merits real agent/CI testing, but it remains a public preview with future metering
Latest change · 2026-10-21：GitHub opens Package Firewall for npm supply-chain controls
The public artifacts support reproduction, while biological validity and publisher performance claims still need independent checks
Latest change · 2026-10-21：NVIDIA releases RFD4-Proteina code and weights
Version 0.5.22 fixes a Mamba-state bug that could silently produce wrong output and expands loading and disaggregated serving; performance still needs reproduction
Latest change · 2026-10-20：Version 0.5.22 expands serving paths and fixes recurrent-state correctness
0.2 adds live Claude Code sessions and an encrypted phone bridge while widening approval and remote-control boundaries
Latest change · 2026-10-18：Littleguys 0.2 adds Claude Code sessions and an encrypted phone bridge
Weights and serving recipes are public, but 1M context and coding results still lack reproducible evidence
Latest change · 2026-10-18：InclusionAI releases Ling-3.1-flash weights for long-horizon agent work
3.0 is GA at the existing price; production sessions on an unpinned alias risk behavioral drift
Latest change · 2026-10-18：xAI makes Grok Voice 3.0 generally available at the existing price
A specialist memory writer has clear engineering value, but evidence still comes from the author-run fixed subsample
Latest change · 2026-10-18：Tencent releases a specialist writer model for long-term agent memory
One-pass structured decisions are worth a trial, but the new checkpoint did not rerun the private board and cannot inherit the older rank
Latest change · 2026-10-18：H2O.ai updates its 4B decision model with two-way reading
CPU-only document parsing is worth a local RAG trial, with performance and license checked on real documents
Latest change · 2026-10-18：Datalab releases a 1M-parameter CPU OCR pipeline
0.1 is bypassable and explicitly not a security gate, so it should not enter mandatory workflows
Latest change · 2026-10-18：Sphinx adds an optional comprehension quiz before Git pushes
Version 0.12.3 adds encrypted collaboration and Redis improvements, while Skill prompt cost and deployment conditions need regression testing
Latest change · 2026-10-17：Open WebUI 0.12.3 adds encrypted collaboration and Redis improvements
The public APIs are low-cost and billed only for successful results, making them useful retrieval baselines; relevance, safety and attribution lack independent validation
Latest change · 2026-10-17：DeepSeek opens Search and Fetch APIs for agent retrieval
Selectable recurrent depth and public weights merit a small trial, but stable upstream runtime support is not ready
Latest change · 2026-10-17：NVIDIA releases Dory1.0-4B with selectable recurrent depth
The public harness and three examples support a practical trial, while the full tasks are controlled and author results remain unreproduced
Latest change · 2026-10-17：FDE-Gym tests agents on underspecified customer delivery
Merge blocking is GA with APIs and audit paths, but it is enterprise-plan gated and false positives and exceptions need validation
Latest change · 2026-10-17：GitHub makes license compliance merge controls generally available
The platform has launched, but SKU stock, pricing and independent performance evidence are inconsistent, so no buying recommendation yet
Latest change · 2026-10-17：NVIDIA launches RTX Spark PCs for local AI development
Open weights and a hosted endpoint are testable, while hardware cost and vendor benchmarks remain constraints
Latest change · 2026-10-16：Beam moves from limited preview to Apache-2.0 open weights
Local weights and a low-cost hosted API are testable, but evidence remains mostly vendor-reported
Latest change · 2026-10-16：Liquid AI launches the hosted d1-ultra decision API
Open weights and small model sizes support a low-cost trial, but leaderboard advantages remain vendor-submitted
Latest change · 2026-10-16：Mistral releases four open-weight Timestral forecasting models
The public model supports validation of kilometer-scale multivariable downscaling, but only inside the X-SHiELD-aligned domain
Latest change · 2026-10-16：Ai2 releases HiRO-ACEv2 for kilometer-scale climate emulation
Open weights and a hosted endpoint make the model testable, while size, gating and custom license terms constrain production use
Latest change · 2026-10-16：Muse Spark moves from an open-weight preview to downloadable and hosted access
Change signals may improve web-discovery efficiency and freshness, but evidence is limited to a research pilot, jointly reported scale and no public interface
Latest change · 2026-10-16：Cloudflare and OpenAI report early results from a web-discovery pilot
Version 0.7.0 improves MoE inference and introduces server compatibility changes
Latest change · 2026-10-16：llama.cpp 0.7.0 adds MoE caching and changes its server API
Temporary registration is effective; catalogue restrictions remain a draft
Latest change · 2026-10-16：China proposes AI export restrictions and starts temporary registration
Gemini 4 Argon has moved from controlled preview to the paid Gemini API tier, enabling direct long-horizon evaluation
Latest change · 2026-10-14：Gemini 4 Argon moves from controlled preview to the paid Gemini API tier and sets a date for its price increase
Maia 300 shows Microsoft's rack-scale inference-efficiency direction, but production deployment is only planned from 2027 and performance claims remain vendor projections
Latest change · 2026-10-14：Microsoft details Maia 300 before production deployments begin
StudentSim provides public code and a personalized AI-tutor evaluation method suitable for rerunning the chess path; simulator results do not replace real learning outcomes
Latest change · 2026-10-14：StudentSim evaluates AI tutors with personalized learner simulations
ColGrep-Agent's open weights and local integrations support an engineering trial, but localization benchmarks and token savings are author-reported and do not yet prove better patch success or total compute
Latest change · 2026-10-14：ColGrep-Agent opens a local model for repository search
Grok 4.7 has a public API, explicit pricing and 500K context for direct comparison; vendor benchmark and safety claims still need independent validation
Latest change · 2026-10-14：xAI releases Grok 4.7 through its public API
MiniCPM-V 4.7 publishes on-device multimodal weights and deployment paths suitable for direct testing; phone speed, quality, thermals and battery use still need independent validation
Latest change · 2026-10-14：OpenBMB releases MiniCPM-V 4.7 for on-device multimodal inference
The paper proves a worst-case blind spot in elicitation-based audits, but its small synthetic construction does not show that deployed models are affected
Latest change · 2026-10-14：Study proves capability audits can miss encrypted model backdoors
GLM-Legal has some independently administered benchmarks but only an application-based beta, with no public weights or generally available API
Latest change · 2026-10-14：Z.ai opens a limited beta for the GLM-Legal model
SSD expert streaming is an engineering idea, but author-only tests and a young runtime do not justify investment now
Latest change · 2026-10-14：Janas streams model experts from SSD on memory-limited computers
The open weights are testable, but the public benchmark is self-scored and not a board entry, so it does not merit investment now
Latest change · 2026-10-14：AutoTrust publishes JEV-35B for one-pass ranking decisions
The launch has community attention but lacks verifiable isolation, reliability and adoption evidence
Latest change · 2026-10-14：Agent37 launches a hosted sandbox service for AI agents
The adaptation is inventive, but an opaque app and author-only metrics do not support engineering adoption
Latest change · 2026-10-14：Independent project adapts EmbeddingGemma 2 for speech transcription
Public weights and broad runtime support make local tool use testable, but vendor benchmarks and the revenue-gated license constrain production adoption
Latest change · 2026-10-14：Liquid AI releases the LFM2.5-24B-A2B reasoning model
The training-free token-saving method has public code, but evidence is author-only and the reference setup needs an 80 GB GPU
Latest change · 2026-10-14：DoubtLess cuts reasoning tokens with contrastive decoding
Weights and inference code are public, making latent-to-tool switching testable; the author results lack independent reproduction
Latest change · 2026-10-14：Hybrid Agent switches between latent reasoning and tool calls
Anthropic-key detection and default push protection are GA, low-cost controls that directly reduce credentials entering Git history
Latest change · 2026-10-14：GitHub push protection now blocks exposed Anthropic API keys
The 10,000-agent long-running system is a strong architecture signal, but the vendor customer story lacks raw telemetry and independent verification
Latest change · 2026-10-14：Third Moment reports running 10,000 agents for multi-day research
Affected snapshots remain callable, but the 2027-01-14 shutdown is fixed, so production users should begin migration validation now
Latest change · 2026-10-14：OpenAI schedules retirement for GPT-4o and GPT-4.1 API snapshots
OpenClaw 2026.10.1 brings permission-aware Skill search, MCP Apps and recoverable backups into stable; search misses, stale saves and migration edges need validation first
Latest change · 2026-10-12：OpenClaw 2026.10.1 adds permission-aware Skill search and recoverable backups
LightOnOCR-3 provides open weights and reviewable materials for document OCR, grounding and chart parsing; vendor benchmarks still need validation on local documents
Latest change · 2026-10-12：LightOn releases OCR models that extract layout, charts and images
Carbon-A publishes the model, data and cross-species benchmark, but broad wet-lab validation on unfamiliar species is still pending
Latest change · 2026-10-12：Hugging Face Bio releases Carbon-A for cross-species gene annotation
Talorys provides a reproducible Cloudflare-hosted personal agent, but its single-user threat model, quota dependence and lack of independent review limit production conclusions
Latest change · 2026-10-12：Talorys packages a self-hosted personal agent for Cloudflare's free tier
Iris-3B publishes pixel-space generation, depth and restoration weights suitable for dense-vision comparisons; author-reported evaluation and high compute cost still need local validation
Latest change · 2026-10-12：Iris-3B explores pixel-space diffusion as a general vision learner
ESP32 iPhone Agent lowers the barrier to real-device iPhone-agent experiments with a low-cost physical bridge, but only author demos exist and cross-app reliability plus safe stopping remain unverified
Latest change · 2026-10-12：ESP32 iPhone Agent turns a low-cost HID board into a phone-control bridge
Agent Standup can locally render Codex and Claude Code sessions for reviewing commands, tool calls and token records
Latest change · 2026-10-12：Agent Standup adds local inspection of Codex sessions
QGIS now explicitly prohibits AI-assisted code in its main repository
Latest change · 2026-10-12：QGIS bans AI-assisted code contributions to its main repository
Agent Lightning v1.0 puts real harnesses directly into RL and opens a lightweight control plane with Kubernetes rollouts
Latest change · 2026-10-10：Microsoft rebuilds Agent Lightning around real agent harnesses
Public endpoints and pricing support a low-cost routing trial, but accuracy, latency and robustness remain vendor-reported
Latest change · 2026-10-10：Microsoft releases a fast model for structured decisions
Community attention is high, but training provenance, license lineage and independent reproduction are insufficient
Latest change · 2026-10-10：A community Qwen derivative spikes without independent validation
Qwen-Image-2.1-Turbo targets lower image-generation latency with eight-step sampling, but lacks a public base-model comparison and uses a custom license
Latest change · 2026-10-09：Qwen releases an eight-step turbo image model
Docker Agent combines YAML agents, multi-agent delegation, MCP, evaluation and OCI distribution in one Apache-2.0 runtime, warranting a small comparison with the existing orchestration stack
Latest change · 2026-10-09：Docker Agent packages multi-agent workflows as OCI artifacts
Whistle provides seven-language transcription, timestamps and speech embeddings in a 16.9 MB CPU model, with public artifacts suitable for target-device testing
Latest change · 2026-10-09：Whistle runs multilingual speech recognition in a 16.9 MB edge model
Public data, code and probe weights support isolated replication, but cross-model transfer, false positives and white-box access cost remain unresolved
Latest change · 2026-10-09：Probes detect hidden agent deception before text monitors do
Millisecond-scale trajectory monitoring has engineering value, but no implementation is public and the abort sample is too small
Latest change · 2026-10-09：OnTrack flags failing agent trajectories in about one millisecond per step
The reanalysis exposes nonlinearity in short tasks but lacks a public replication package and independent response
Latest change · 2026-10-09：Study finds equal-looking AI time-horizon gains can hide unequal difficulty
Public 5B weights cover many inputs, but evaluation code is not released and vendor benchmarks plus the custom license need review
Latest change · 2026-10-09：Tencent releases one parser for documents, charts, audio and video
Thirty days of production data make the metrics useful, but the scheduler and simulator are not public, so only the contracts are reusable
Latest change · 2026-10-09：Ai2 reallocates scarce GPUs with budget-weighted fair sharing
The three-tier CVP binds advanced cyber capabilities to verification, retention and monitoring controls
Latest change · 2026-10-08：Anthropic expands cyber access into three verified tiers
Claude Haiku 5.5 is a mature, multi-cloud low-cost long-context model for high-volume and subagent routing pilots
Latest change · 2026-10-08：Anthropic releases Claude Haiku 5.5 for high-volume work
EMA Lightning is a small, runnable open Turkish streaming TTS model
Latest change · 2026-10-08：EMA Lightning releases a compact streaming Turkish speech model
EmbeddingGemma 2 is an Apache 2.0 multimodal embedding model suitable for on-device retrieval and RAG pilots
Latest change · 2026-10-08：Google releases EmbeddingGemma 2 for on-device multimodal retrieval
openTPU is an inspectable FPGA inference learning and validation platform with a low-cost simulator path; it is not a mature production accelerator
Latest change · 2026-10-08：openTPU runs open models on a fully inspectable FPGA accelerator
ScienceClaw-Eval uses replay and independent-task gates to evaluate agent self-improvement, with public artifacts for isolated reproduction
Latest change · 2026-10-08：ScienceClaw tests whether scientific agents retain verified improvements
Fixed opening tokens can materially change base-model reasoning and refusal behavior, but effects may depend on model, decoding and task
Latest change · 2026-10-07：Token cues unlock latent reasoning behavior in base models
MemPilot turns multimodal memory curation into a runtime quality, cost and latency routing problem; public code supports a bounded trial
Latest change · 2026-10-07：MemPilot learns when and how an agent should curate multimodal memory
BazaarBench provides a public marketplace-agent safety stack that checks inventory, commitments and transaction consequences
Latest change · 2026-10-07：BazaarBench exposes double-selling failures in marketplace agents
OPPD may distill multi-candidate reasoning search into one generation, but author results need reproduction with full compute accounting
Latest change · 2026-10-07：OPPD distills expensive answer search into one model generation
Falcon-Emirati shows the value of regional dialect models, but chat-only access blocks independent deployment and verification
Latest change · 2026-10-07：TII tunes Falcon-Emirati-7B for Emirati dialect and culture
Agents produced an auditable materials hypothesis, but key properties lack experimental confirmation
Latest change · 2026-10-07：Claude agents propose two room-temperature magnetic semiconductor candidates
A verified externality incident shows production agents need identity, egress, rate and attributable audit controls
Latest change · 2026-10-07：Wikimedia finds unauthorized OpenAI agents editing and probing its services
Dust publishes a minimal transformer training implementation without backprop, but remains orders of magnitude from production efficiency
Latest change · 2026-10-07：Dust trains small transformers without backpropagation
textGrain has entered API opt-in and phased EU product rollout, but short text, rewriting attacks and a restricted detector limit independent validation
Latest change · 2026-10-06：OpenAI launches opt-in textGrain watermarking for API output
Downloads are high, but reproducible and independent evaluation is insufficient, so the models do not merit investment yet
Latest change · 2026-10-06：AutoTrust decision models surge on Hugging Face without independent validation
Sonnet 5.5 task cost is effort-sensitive; independent evaluation found maximum effort can cost more than Sonnet 5
Latest change · 2026-10-05：Independent evaluation finds Sonnet 5.5 maximum effort can cost more
The Agents Vault API is publicly documented; secrets are injected by a network proxy per allowed host, enabling a controlled trial
Latest change · 2026-10-05：OpenAI publishes Agents Vault API docs, clearing the trial gate
ThinkingBox directly grades terminal state and side effects; the public harness is ready for repeatability trials
Latest change · 2026-10-05：Microsoft opens ThinkingBox to check agents against final database state
DiarizationLM ships runnable local weights for speaker-label correction, but meeting evidence remains small
Latest change · 2026-10-05：Google opens a 4B model that repairs speaker labels in transcripts
Strata merits a trial for adoption and heterogeneous serving, while speed and low-bit quality need independent checks
Latest change · 2026-10-05：Strata runs Qwen3.8-Flash-Next on 12 GB gaming GPUs
Ataraxos has a paper and training code but no checkpoint, so it remains watch-only
Latest change · 2026-10-05：Ataraxos beats Stratego champions with damped self-play and search
Figma hosted MCP uses a client allowlist, so catalog eligibility must be checked before integration
Latest change · 2026-10-05：Figma limits its hosted MCP server to catalog-listed clients
Strands Decider is a small decision model with public weights, training code and evaluation records, suitable for agent routing and tool-selection trials; author results still need reproduction on real distributions
Latest change · 2026-10-04：Strands releases a 2B open decision model for agent control
MAI Speech offers callable low-latency streaming transcription and synthesis with published pricing; multilingual quality and interruption recovery on real calls still need independent validation
Latest change · 2026-10-04：Microsoft releases streaming speech models with 100 ms partials
Index-Translate provides Apache-2.0 preview weights and inference code for direct tests of terminology and structured-output constraints; vendor-only benchmarks and preview status do not support production use
Latest change · 2026-10-04：IndexTeam opens a 35B MoE model for constrained translation
UI-Venus-2's 9B weights and framework are public for isolated cross-interface evaluation, but the weight license remains pending and blocks commercial deployment
Latest change · 2026-10-04：The 9B weights and framework are public, but the license remains pending; the rating moves from WATCH to TRIAL
Kolibri provides open weights and a serving plugin for German-English RAG, tool calling and long documents; vendor evaluations still need independent reproduction
Latest change · 2026-10-04：Aleph Alpha opens Kolibri for German-English reasoning
Naive-N0.5-Flash is an open sparse model with native 1M context, but roughly 315 GB of weights and vendor benchmarks limit trials to FP8 clusters
Latest change · 2026-10-04：NaiveAI opens a 309B sparse model with native 1M context
ds4 offers a specialized local runtime and strong adoption signal for a few large MoEs on high-memory workstations, but model coverage is narrow and the software remains beta
Latest change · 2026-10-04：DwarfStar builds a narrow local runtime for frontier open weights
AstaBrief-8B is a downloadable, permissively licensed cited-report model, but citation quality and the small human study still need independent verification
Latest change · 2026-10-03：Ai2 opens AstaBrief-8B for fast cited research reports
PixelUMM's unified pixel-space architecture is inspectable, but its checkpoint is noncommercial and both the training recipe and production evidence are incomplete
Latest change · 2026-10-03：PixelUMM opens a single pixel-space model for understanding and generation
cua-speedrun is a public, standardized stack for measuring computer-use agent speed, cost and success
Latest change · 2026-10-02：cua-speedrun standardizes speed and cost tests for computer-use agents
Olmo-core 3 is a rare open reference stack for trillion-parameter MoE training, with inspectable code and design
Latest change · 2026-10-02：Ai2 opens Olmo-core 3 for trillion-parameter MoE training
Dots productize persistent execution, app connections and individual cloud computers, but access, later pricing and enterprise governance remain unsettled
Latest change · 2026-10-02：OpenAI launches Dots as always-on cloud agents
Kumo Tabular is an open-weight tabular foundation model that reads labeled rows as context and classifies or regresses in one forward pass
Latest change · 2026-09-30：NVIDIA releases Kumo Tabular for zero-training table prediction
Holo4 ships weights, a harness and step-by-step trajectories together, and one model can operate GUIs, run code and call MCP or APIs
Latest change · 2026-09-29：H Company releases Holo4 to work across screens, code and APIs
Tencent Hy4 preview is a very large Apache 2.0 MoE with a 1M context, and its main repository has reached 22,788 monthly downloads
Latest change · 2026-09-28：Hy4 preview reached 22,788 monthly downloads, but the full model and reproducible independent evaluation remain absent
Pydantic AI fixed a web-fetch policy bypass and telemetry content leaks, which are runtime-boundary flaws rather than cosmetic patches
Latest change · 2026-09-27：Pydantic AI 2.51.0 adds GPT-Live support and realtime context usage while reducing agent graph overhead
Microsoft is turning Copilot into a tenant-governed agent workspace: Code builds, Managed Runtime handles sandboxing and hosting, and Autopilot runs recurring work
Latest change · 2026-09-27：Microsoft turns Copilot into a tenant-governed agent workspace
Cursor is extending its coding agent from the editor to always-on cloud execution: self-hosted workers, Projects for months-long work, deployment monitoring and security review have all shipped
Latest change · 2026-09-25：Cursor extends coding agents into deployment monitoring and security review
GLM-5.3 and GLM-5.3-Flash carry the most permissive license (MIT) in the current wave of Chinese frontier open weights, and Flash uses hybrid sparse-linear attention to cut long-context serving cost
Latest change · 2026-09-12：A one-direction weight edit strips refusal from the shipped GLM-5.3-Flash weights
The archive has no release record for the Kimi models themselves, only vLLM 0.28 adding model-specific serving work for Kimi-K3
Latest change · 2026-08-30：vLLM 0.28.0 lands a big-model serving push for Kimi-K3 and DeepSeek V4