The Hub for AI Coding Models, Tools & Workflows

Shanghai AI Lab’s LMDeploy shipped v0.18.0 on 2026-09-28: input logprobs, expanded SM90 quantized GEMMs with unified TurboMind linear paths, Ascend GLM-5.2, Qwen3.5 dflash, and MoE shared-expert/FFN sharding. PyTorch path prototypes TurboMind W4A16 (AWQ), piecewise CUDA Graph prefill, XTuner TileLang sparse MLA, checkpoint-engine weight updates, and request-only KV cache metrics.

MCP’s TypeScript monorepo shipped v2.2.0 on 2026-09-28: client/server/core/server-legacy/codemod aligned at 2.2.0. M2M OAuth providers should pass `expectedIssuer`; `fetchToken()` throws `AuthorizationServerMismatchError` on issuer mismatch. Cursor-less listTools/listPrompts/listResources/listResourceTemplates now follow `nextCursor` (capped by `listMaxPages`). Fixes the 2.1.0 CJS jose types regression and Client.listen() unhandled rejection / send hang.

Anthropic shipped Python SDK v1.9.0 on 2026-09-28: adds claude-sonnet-5-5, between_tools thinking, GA cache diagnostics on Message/MessageCreateParams, and optional tool execution while the reply streams. Managed Agents event filters and workspace rate-limit include_inherited/source land too, plus unnamed upload and stream()/parse() diagnostics fixes.

LangChain4j 1.20.2 (2026-09-28) focuses on Agent-to-Agent (A2A) client fixes: `tenant()` on `A2AClientInstance`/`A2AClientBuilder`, `@A2AClientAgent(tenant=…)` auto-attaches tenant on every `MessageSendParams` (or parses it from `/.well-known/{tenant}/agent-card.json`), and filters tenant method args so they are not exposed as LLM tool parameters. AgenticScope/ResultWithAgenticScope toString guards against circular refs.

Unsloth shipped v0.1.900-beta on 2026-09-28 (PyPI unsloth 2026.9.12): Desktop runs Laya Decision models locally via a Jev-compatible `/v1/systemone` endpoint, adds a Skills CRUD editor and Library viewers for PDF/Office, and improves Apple Silicon with batched serving, structured outputs, and TurboQuant KV. Media pipelines claim ~4.5× LTX-2.3 clips, 1.7–6.3× VAE decode, and up to ~1 minute faster MiniMax-H3 first render with 25–29 GiB lower peak VRAM.

OpenAI shipped Python SDK v3.20.0 on 2026-09-28: Agents credentials gain metadata and update-without-replace auth, sessions add ultrafast tier plus misalignment_policy_violation; Responses can select Cyber access programs (standard/daybreak_blue/daybreak_red); and an opt-in ResponsesWebSocketAccumulator yields immutable incremental text/tool snapshots. Live/Realtime WebSocket query/queue/TLS retry fixes land in the same cut.

Firecrawl published firecrawl-py 4.45.0 and @mendable/firecrawl-js 4.42.0 on 2026-09-28, shipping `agent_hints` preservation from [#4641](https://github.com/firecrawl/firecrawl/pull/4641): with `X-Firecrawl-Agent-Hints: true`, Search/Scrape/Parse/Map responses can include deterministic next-step suggestions. Also includes Hangar browser/interact migration and `exchange.onTermsRequired` fields.

OpenCode v1.18.33 (2026-09-28) makes Cloudflare AI Gateway models honor provider response/stream timeouts, reports MCP browser launcher exits, redacts credentials in debug config output, and aligns Gemini thinking defaults/effort options with supported controls across model generations.

Docker shipped cagent (docker-agent) v1.145.0 on 2026-09-28: ACP requests now propagate W3C `traceparent`/`tracestate` with server spans for each of the 13 implemented protocol handlers, plus new lint cops and a lean-TUI fix that preserves scrollback during partial streaming tool calls.

While looped language models enable depth-adaptive inference by executing fewer layer repetitions on easy tokens and more on complex tokens, variable loop counts break conventional batching engines like vLLM. Researchers from TUM and Imperial College introduced Continuous Depth Batching (CDB). By dynamically reorganizing batches between loop iterations, orchestrating looped KV-caches, and asynchronously predicting token exits, CDB achieves up to 99% of the theoretical maximum inference speedup across Ouro 1.4B and Huginn 3.5B architectures.

Shanghai AI Lab and OpenDataLab open-sourced InternW0-Δ, a unified World Action Model (WAM) pretrained on over 20,000 hours of curated robotic demonstrations, UMI data, and egocentric videos. Built on a Mixture-of-Transformers (MoT) architecture that binds visual predictive dynamics with robot actions, it introduces 'Causal Imprint' to inject future representations directly into the action expert without requiring costly future-video rollouts during inference.

Addressing the rapid saturation and benchmark overfitting of static LLM evaluations, Google DeepMind and Kaggle unveiled Game Arena. By orchestrating head-to-head live agent matchups with dynamic Elo rating evolution, Game Arena prevents performance saturation across three pilot environments: Chess (perfect information), Texas Hold'em Poker (imperfect information), and Werewolf (multiplayer social deduction and deception), systematically benchmarking long-horizon strategic reasoning.

Google shipped ADK Python v2.10.0 on 2026-09-25 (`google-adk` 2.10.0): experimental skill lifecycle (`ADK_ENABLE_SKILL_LIFECYCLE=1`) with ephemeral lifecycles and active skill limits; MongoDB toolset for vector/hybrid search; evaluation metrics for duration, tokens, and model-call counts; better OpenAI reasoning model parameter adaptation and reasoning-token reporting. BigQuery protected write mode and instruction templating tighten with breaking changes.

BerriAI shipped LiteLLM v1.103.0 on 2026-09-28: delegated MCP OAuth now requires admission; Capability classifier + Fuse V2 routers land; the proxy can predict prompt-cache costs across deployments, bind JWT claims to agents, add `tpd_limit`, bulk user/team APIs, and `/nvidia_nim` passthrough. Also: Bedrock S3 managed file delete/list, Friendli price auto-sync, Responses↔Chat reasoning mapping, and cosign image verification docs.

OpenAI shipped Codex rust-v0.158.0 on 2026-09-28 (`@openai/codex` 0.158.0): MCP servers with pre-registered OAuth client secrets (`codex mcp add --oauth-client-secret`); bearer-token auth for exec-server WebSockets; fullscreen TUI copy-on-select/right-click paste with Markdown-preserving copies; explicit transparent backgrounds for image gen/edit. Elevated-permission terminal approvals default on; Windows/Linux/macOS sandbox and approval-retry fixes land too.

LiteLLM shipped v1.103.0, introducing native hosted_vllm offline batch processing and updating its MCP integration test suite to the modern MCP 2.x MCPServer architecture. The Admin UI now supports dynamic web search interception controls, while the proxy resolves reasoning-object to reasoning_effort translation for OpenAI o1/o3 series and eliminates Azure 400 errors on empty tool choices.

Challenging the assumption that agent memory must rely solely on external explicit retrieval tables, researchers unveiled MemoryAthena. By establishing three memory pathways—direct Engram retrieval (E), cue-conditioned generation (GE), and causal backbone generation (GH)—governed by a lightweight 201M causal routing head trained on counterfactual token likelihood advantages, MemoryAthena raises five-task QA accuracy from 37.65 to 39.28 while boosting general NLP averages to 79.13.

Addressing the systemic bottlenecks in multimodal document processing (MinerU, Docling) where variable-length child generation breaks GPU batching and lineage tracking, Peking University and OpenDCAI open-sourced RayOrch. Operating under a lineage-controlled multi-grain dataflow model with per-call FIFO ready queues, RayOrch delivers a 15.14x speedup scaling MinerU from 4 to 64 NVIDIA H20 GPUs, cutting end-to-end execution time by 13.1%~16.0% compared to Ray Data and 29.0% compared to Daft.

Researchers from SJTU and GAIR introduced PISA, a block-sparse attention mechanism that eliminates the quadratic bottleneck of conventional block selection in long-context LLMs. By adopting a coarse-to-fine pyramid Top-K selection hierarchy across O(log N) key levels, PISA reduces complexity to O(N log N). Accompanied by hardware-aware fused Triton kernels that avoid materializing query-key matrices, it matches dense baselines while substantially boosting long-context retrieval throughput.

Overcoming the limits of traditional multi-agent benchmarks that only test short-horizon (<20 steps) or competitive settings, UPenn researchers released AgentWorld. Spanning 50+ interaction rounds in an MMORPG sandbox across 100 human-annotated tasks with 3-20 asymmetric black-box agents, it introduces Causal Collaboration Effectiveness (CCE). Even frontier models achieve only 52.0% task success, exposing severe communication breakdown and shared plan decay.