Liquid AI releases LFM2.5-24B-A2B, a 24B-parameter reasoning model with 2B active parameters for on-device agents. The company says it leads or matches Qwen3.6-35B-A3B on five of eight agentic benchmarks, scoring 27.69 on AutomationBench versus 8.72. The open-weight model decodes 184 tokens/s on an Apple M5 Max.
NVIDIA releases the full Run:ai platform as open source software inside its DSX AI Factory Platform, including KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map. The company says KAI Scheduler ships under Apache 2.0 via CNCF, and GPU Fractioning Operator uses CUDA 13.4 native fractioning. Crusoe, Mirantis, Nebius, and Rafay are among eight partners integrating it.
BichengWang publishes three MIT-licensed Python packages, jev-decision, decision-control and experiment-gate, that emit replayable records for AI go/no-go decisions. They run on Python 3.11+ using only the standard library, and sit on GitHub at version 0.1.0 with PyPI publication still pending. A fourth tool, RELIA, stays unreleased and human-gated.
Elorian and Scale AI release Humanity's Sixth Sense, a 522-question visual reasoning benchmark, where GPT 6 Astra scores 53.6 pass@1 against a human 93.1. Reflection says its Beam model, 501 billion parameters with 23 billion active, trails Chinese open weight rivals on benchmarks.
Fish Audio releases Drama 3, an emotive text-to-speech model it says tops Artificial Analysis' Pronunciation Robustness benchmark at 93.2%, against 88.1% for its S2.1 Pro. The API costs $10 per million characters through November 12, 80% below the $50 list price. Open weights carry a $500,000 revenue cap for commercial use.
Mistral AI opened Mistral Large 4 to public preview on October 6, 2026, an open-weight multimodal model with 52 billion active parameters, 1.05 trillion total, and a 1.6 billion parameter vision encoder. Mistral lists a 1M token context and prices input at $0.68 per million tokens and output at $2.09.
Sarthak Rastogi releases agent-engineer, an open-source plugin for Claude Code, Codex CLI, Cursor, OpenCode and Gemini CLI that adds 11 skills, 4 subagents and 5 hooks for agent work. Rastogi reports that in a blinded six-task test, Claude Code (Opus 5.5) met 53 percent of requirements with the plugin versus 34 percent without.
Google Agent Skills packages Google Cloud domain knowledge into open-source instructions for AI coding agents, lead developer relations engineer Remigiusz Samborski says. Skills pass automated linters, link checks, and on-submit plus weekly evaluations. Google reports the project drew over 15,000 GitHub stars since its Google Cloud Next 2026 launch, and says skills should prefer remote MCP tools.
SemiAnalysis finds that of 857 model releases from nine Chinese labs between September 2021 and September 2026, only 31, about 3.6%, carried safety reports. Alibaba shipped 238 models with seven reports. The firm says enterprises adopting cheaper Chinese open-source models for cost and control take on the safety burden themselves.
AirLLM's v4.0.0 release adds training support for Qwen3.8-Flash-Next, a 125B model, under 6GB of VRAM, streaming frozen weights one decoder layer at a time while adapters stay on the GPU. The project, maintained by Gavin Li, says it also runs Kimi K3 (2.8T) in 3.72GB. The Apache-2.0 library installs via pip.