276B total / 12B active (MoE)
1M tokens
Apache 2.0
Efficient open-weight multimodal reasoning model that delivers performance comparable to the 975B-total, 41B-active Inkling flagship at roughly one quarter of its active and total parameter counts. Inkling-Small is a 42-layer decoder-only autoregressive transformer with a sparse Mixture-of-Experts feed-forward backbone: each token is routed to 6 of 256 experts alongside 2 shared experts, while a hybrid of local and global attention layers supports context windows of up to 1M tokens. It uses the same natively multimodal, encoder-free design as Inkling, jointly processing text with image representations produced by a four-layer hierarchical patch encoder and audio represented through discrete dMel spectrogram tokens; accepted inputs are text, images, and audio, with text output. The final model follows an earlier preview checkpoint and incorporates an improved pretraining data mix and training recipe, on-policy distillation from Inkling, and two additional weeks of scaled agentic-coding reinforcement learning. It supports variable thinking effort so deployments can trade reasoning compute and output length against performance. Thinking Machines reports that the final release surpasses Inkling on several reasoning and agentic-coding evaluations while retaining competitive multimodal performance, including 80.2% on SWE-Bench Verified, 55.9% on SWE-Bench Pro Public, 64.7% on Terminal-Bench 2.1, 89.5% on GPQA Diamond, 31.6% on Humanity's Last Exam text-only, 47.8% on Humanity's Last Exam with tools, 95.5% on AIME 2026, 40.1% on ARC-AGI-2, 74.0% on MMMU Pro, 54.9% on Audio MC, and 90.1% on VoiceBench at its published high-effort evaluation setting. The full checkpoint is available in BF16 and an official NVFP4 quantization, with documented local deployment support through SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face Transformers.