Thinking Machines launches Inkling 2 Small: 594B-param MoE, free on OpenRouter
thinkymachines · x · 2026-11-10
Thinking Machines Lab released Inkling 2 Small, free for a limited time on OpenRouter.
- Architecture: multimodal mixture-of-experts reasoning model with 17B active params out of 594B total; first model of the second Inkling generation
- Training: RL on long-horizon coding tasks across many agent harnesses, targeting agentic coding, autonomous research, knowledge work and RAG
- Specs: text + image input (no audio), tool calling, variable reasoning effort, 1M-token context window
- Caveats: the free endpoint is restricted to agentic harnesses like Claude Code, Codex and Hermes Agent; prompts and outputs are logged (de-identified) to improve TML's models — don't upload confidential data
Notable hook: it pairs vision with agentic coding, using visual feedback to build and refine websites, games and 3D designs.
Related event: Thinking Machines Launches Inkling 2 Small on OpenRouter(2 posts)→
More from coding & agent
- Physical-Engineering's Last Exam launches on Kaggle with $400K in prizes — MeganRisdal · 2026-11-10
- Elastic TTS: custom text-to-speech engine coded by Gemini 4 Argon — alexanderchen · 2026-11-10
- 'Proof-of-Agent' mining: one user runs POA across Hermes, OpenClaw, Grok Bot and Muse — adamamcbride · 2026-11-10
- Vercel MCP opens to all clients via Client ID Metadata Documents, ending allowlist friction — steren · 2026-11-10
- User Lets Fable 5.2 Max 'Do Whatever It Wants' — It Keeps Working for Over an Hour — mallow610 · 2026-11-10
- Simon Willison vibe-codes a Beep Locator with Claude Code to triangulate a mystery beep using five devices — RealGeneKim · 2026-11-10