Inkling-Small: 276B/12B open weights beat 975B Inkling on SWE-Bench
Thinking Machines shipped the full Apache-2.0 weights: 1M context, native audio+image, 80.2% SWE-Bench, day-0 GGUFs, and fine-tuning on Tinker today.
Copy markdown
Quarter the size, better at code
Thinking Machines released the full open weights for Inkling-Small: 276B total but just 12B active per token, versus the flagship Inkling's 975B/41B. It hits 80.2% on SWE-Bench Verified — edging its 4x-larger sibling's 77.6% and making it a cheaper model that's actually stronger at agentic coding.
Apache 2.0, 1M context, natively multimodal
The weights ship under Apache 2.0, so commercial use is fair game. You get up to 1M-token context plus native image (to 4096px) and audio (16kHz WAV) input with no bolt-on adapters, and reasoning scores near closed frontier models: 89.5% GPQA Diamond and 95.5% AIME 2026.
Run it yourself: day-0 Unsloth GGUFs
Unsloth posted GGUF quants at launch — roughly 75GB (1-bit) up to 197GB (5-bit) — servable through llama.cpp or Ollama as an OpenAI-compatible endpoint. It's a workstation model, not a laptop one, but only 12B active params keep tokens/sec high even with CPU offload.
Fine-tune it today on Tinker
You can fine-tune Inkling-Small right now via Thinking Machines' Tinker API, or try it in text, image, and audio on Tinker Playground first. With Apache-2.0 weights, that's a full own-and-customize path rather than renting a closed endpoint.
Build this weekend
An Apache-2.0 model that ingests audio and images and fine-tunes cheaply unlocks a self-hosted multimodal agent with no per-token bill — think a voice-driven coding assistant, or a private 'screenshot + mic to PR' tool, with 1M context for whole-repo passes.