Thinking Machines releases Inkling, an open-weights model built for tuning

Thinking Machines Lab released Inkling, its first open-weights model, and the framing is deliberate: the team calls it "a good open-weights base for customization," not the strongest model you can run. Inkling is a 975B-parameter mixture-of-experts with 41B active, alongside a smaller 276B/12B preview, and both handle context windows up to 1M tokens. The full weights are on Hugging Face, including an NVFP4 checkpoint for NVIDIA Blackwell, with fine-tuning offered at half price on the company's Tinker platform.

The architecture has some unusual choices worth reading if you tune models. It uses relative positional embeddings instead of RoPE for better long-context extrapolation, interleaves sliding-window and global attention at a 5:1 ratio, and adds short convolutions after the attention key and value projections. The multimodal path is encoder-free: audio arrives as dMel spectrograms and images as 40x40 pixel patches, fed straight into the model. Pretraining ran on 45 trillion tokens of text, images, audio, and video, and post-training used more than 30 million reinforcement-learning rollouts.

Benchmarks land in competitive open-weights territory rather than at the frontier: 77.6% on SWE-bench Verified, 97.1% on AIME, 87.2% on GPQA Diamond. The more practical feature is controllable thinking effort. At lower settings Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly a third of the tokens, and the team says its chain-of-thought grew more compressed on its own during RL training, without being optimized for brevity. The bet is breadth across domains and cheap adaptation over a single headline score.

Why it matters

If you fine-tune open models, Inkling is aimed squarely at you: a broad multimodal base with a permissive release, discounted tuning, and a token-cost dial that lets you trade reasoning depth for cost per task. Worth testing against the open MoE you use now, especially if your workload is domain-specific rather than leaderboard-shaped.