Explore complete model, tool, agent, skill and harness dossiers and related events.
OpenAI announced GPT-Live-1.1 and Quip on 10/15, following Live-1 and mini. Full-duplex interaction listens, responds and handles interruptions continuously, with thinking-model delegation. The source is a system card, not verification of a public API identifier or price.
Anthropic's model family. Fable, Opus, Sonnet and Haiku retain separate versions, access conditions, prices and migration materials. This update adds Haiku 5.5 and the Sonnet cache-read price reduction.
Whistle provides CPU speech recognition on Needle's C++ runtime. Its model file is 16.9MB, with word timestamps, confidence and keyword biasing. Officially listed languages are English, German, French, Spanish, Italian, Dutch and Polish; Chinese is not listed.
Alibaba's Qwen model family. Dense models, experimental sparse architectures and audio/video understanding branches retain separate versions, access methods and evidence.
Aleph Alpha's Kolibri-1 provides Apache-2.0 weights, with reasoning, tool calls and long-document processing in German and English.
Ai2's Qwen3-8B-based AstaBrief takes research questions and paper excerpts to generate cited reports with a smaller model. Its model card specifies Apache-2.0.
Cloudflare's Clef and Clef-flash provide decisions and probabilities over predefined types and options, with Workers AI and open-weight access.
NVIDIA's Kumo Tabular foundation model has three sizes from 28M to 215M parameters. Labeled rows supply context for class probabilities or regression outputs on new rows.
A structured decision model with open weights and a Python SDK: provide state and a question definition to receive choices, scores or yes/no probabilities. It supports research into fixed-output classification and routing workflows. This independent project resembles Jev in function; FieldToFit has not verified it as an equivalent replacement.
Google announced broader access on 10/13, following the 9/30 Fairwind partner phase. This changes access conditions rather than the original launch date.
Moonshot AI's model family, covering native vision models and a specialized coding branch.
Zhipu AI's model family. This profile distinguishes the GLM-5 main line from the Flash branch using a new base model.
DeepSeek's model family. This profile focuses on the V4 Flash branch's inference architecture and multimodal changes.
MiniMax's model family. Initial coverage details the general-purpose and agent-model main line; music and other generation branches will have separate tables later.
ByteDance Seed's model research and releases. Doubao product names, research names and API model identifiers remain separate.
Durable Actors provides persistent state and serialized execution for group chat, collaborative applications and agent coordination. Version 0.8.0 was released at 05:51 Beijing time on 10/11, adding native Python development tools, separate runtime images and storage integrity checks.
GitHub published the release at 20:32 Beijing time on 10/14; its changelog heading says 10/13. This follows the existing REA6.4.0 dossier rather than resetting the original discovery.
Cloudflare's Web Search API beta connects Ceramic.ai, Exa and Linkup through AI Gateway, returning titles, URLs and descriptions.
Released at 03:26 Beijing time on 10/14, adding Unsloth quantization choices, NVFP4 packages, Qwen3.6-35B-A3B support and an experimental Metal engine.
AG-UI 1.0 provides a stable event specification between agents and application interfaces, covering subagents, human interruptions, multimodal tool results and token usage.
Meta's Muse Gadgets provides ESP32 and Linux device SDKs, connecting Muse to screens, buttons, sensors and Raspberry Pi programs.
Ai2 describes the Olmo-core 3 training system, focusing on communication, expert parallelism and throughput overhead when scaling MoE. The article's “3” does not directly identify a release tag of the stored code package.
Data Agent Kit is generally available. It combines MCP tools with service-specific skills to help coding agents work with cloud-data workflows.
Cloudflare updates the Workers-based Kitesurf browser with WebMCP, more web-standard support, Quick Actions calls inside Workers and terminal-page rendering.
Hugging Face's open-source relore turns repository issues, PRs, comments and current code into searchable project memory, helping coding agents find prior work and decisions.
Whiteboard is an open-source desktop collaboration canvas that connects existing coding agents, linking architecture diagrams, design explanations and decision records to local code.
Released at 11:07 Beijing time on 10/13, adding Liquid AI d1 text/image decision paths, Decima-agent 2.2 and retrieval by manifest digest.
Released at 05:29 Beijing time on 10/14, adding asset export, standard 3D output, camera nodes and native LoRA-stack loading.
Run models through a unified command-line and service interface, connecting model files with the applications that use them.
An open-source inference and serving engine that continually adapts to model architectures, numerical precision and hardware.
Released at 20:18 Beijing time on 10/14, adding per-page ConfidenceScores, threaded-backend concurrent-result controls, progress and event callbacks.
An MCP testing tool that lets AI operate physical or emulated Android devices, with screenshots, logs and execution replays.
Lookout was released at 02:10 Beijing time on 10/14. nanoMuse is a community personal agent, unaffiliated with Meta’s official Muse.
Released at 16:21 Beijing time on 10/15, 0.22.2 discloses that 0.22.1 was built from the wrong commit and included unreleased Gateway work, leaving some Microsoft Store installs at Connecting. FieldToFit adds a correction to the earlier event.
OpenAI's coding agent, working with project files, tool execution and ongoing tasks.
The CLI and Python SDK were released at 03:22 and 04:22 Beijing time on 10/16; GitHub’s UTC date is 10/15. This maintains waiting, worktree protection and error reporting rather than models or usage limits.
The stable release arrived at 04:44 Beijing time on 10/14. This is client maintenance rather than a Gemini model-capability or pricing update.
An open-source ecosystem for software-development agents. Its current main repository provides Agent Canvas, a console for local, remote and cloud agents.
Released at 07:50 Beijing time on 10/16, with a GitHub UTC date of 10/15, 0.13.13 adds an OpenAI browser-tool driver for local, Browser Use Cloud and borrowed browser sessions.
Released at 08:36 Beijing time on 10/16, v2026.10.4 is a stability patch after 10.3. The dossier previously recorded v2026.9.5; this checks specific 10.4 fixes while intermediate releases await review.
Pi is an agent CLI, SDK and extension framework. Version 1.2.0 was released at 07:03 Beijing time on 2026-10-13, adding per-key model lists, configuration schemas and extension-event compatibility changes.
Released at 04:21 Beijing time on 10/14, following 0.6.12. Reading these materials does not install, load or execute the skills.
A coding-agent workflow for requirements clarification, planning, testing, debugging and acceptance checks.
Anthropic's public skill collection shows how task instructions, scripts and materials form loadable skills.
Vercel's skill collection packages frontend-engineering and website-operation experience for coding agents.
An open-source skill for AI-generated technical diagrams. It chooses diagrams for system relationships, execution flows or interaction sequences, produces editable HTML/SVG, and documents SVG/PNG export. Useful for technical articles, project documentation and paper-method explanations; the user's own AI interprets and generates the content.
BootLoops combines mathematical-physics computation tools with agent working conventions for research requiring exact or high-precision results, including integrals, recurrences and Bayesian evidence.
AgentRun defines steps, branches and bounded loops in JSON or TypeScript. It calls Jev for explicit decisions and existing agents for investigation, while the host manages tools, permissions and budgets.
A general-purpose agent harness assembled from Strands Agents, with default tools, context, sessions and memory. Use it directly or recombine its components through the SDK.
An agent runtime based on the Cordis plugin system. Plugins compose models, tools, sessions and execution loops.
The core SDK was released at 09:13 Beijing time on 10/14. Filesystem rules and interrupt_on approvals check both requested paths and their resolved targets.
The CLI and Python SDK were released at 03:22 and 04:22 Beijing time on 10/16; GitHub’s UTC date is 10/15. This maintains waiting, worktree protection and error reporting rather than models or usage limits.
An agent SDK and service interface for software tasks, organizing tools, conversations, workspaces and execution events.