Thinking Machines Releases Inkling-Small With Apache 2.0 Weights
The 276-billion-parameter mixture-of-experts model activates 12 billion parameters per token and ships for self-hosting, fine-tuning, and multimodal use.
Thinking Machines Lab released Inkling-Small on July 30 as an Apache 2.0 open-weights model for developers building agentic systems, coding assistants, chatbots and other multimodal applications. The model has 276 billion total parameters, activates 12 billion per token and supports a context window of up to one million tokens, according to the lab's release article and model card.
What changed
The finished release turns the earlier Inkling-Small preview into a separately distributed model with downloadable weights. Thinking Machines also makes it available through Tinker for fine-tuning and for text, image and audio chat. The model card describes a decoder-only mixture-of-experts transformer whose modalities share one processing space, with text output from text, image or audio inputs.
Thinking Machines positions Inkling-Small as a more compute-efficient counterpart to Inkling. The company says it reaches comparable overall performance at roughly one quarter of Inkling's size. It also says changes to the training data mix, pre-training recipe, distillation and two additional weeks of agentic coding reinforcement learning helped the smaller model surpass Inkling on reasoning and agentic coding evaluations. The company still gives the larger model an advantage in knowledge coverage and factuality.
Why it matters
The release gives teams an open model that can be owned, modified and integrated into third-party products instead of being limited to a hosted chat or closed API. That flexibility does not make self-hosting lightweight. Thinking Machines says the BF16 checkpoint needs at least 600 GB of aggregate GPU memory. A quantized NVFP4 checkpoint lowers the stated minimum to 180 GB, which still places direct deployment in workstation-cluster or data-centre territory for most users.
The model's multimodal scope also broadens the kinds of local or customized systems developers can test. The official intended-use list includes tool-using agents, coding assistants, retrieval-augmented generation and conversational applications across multiple programming and spoken languages.
What remains uncertain
Performance comparisons should be read as vendor claims. Thinking Machines uses a mix of its own evaluations, externally reported results and self-reported figures for other models, with harness-specific caveats on several coding tests. This run did not find an independent evaluation of the finalized weights.
The model card also warns that Inkling-Small can hallucinate, miss instructions, degrade in long conversations and perform unevenly across languages or domains. It reports residual cases where indirectly framed or role-play prompts can elicit harmful-topic compliance, so the company recommends application-layer filtering, monitoring and human oversight for sensitive uses.
Status
Confirmed. The model, weights, license, distribution options and model card are published by Thinking Machines Lab. Internal confidence is medium because capability and safety claims come from the model provider.
Sources
Update note: Last reviewed 2026-08-03. We will revise this post if independent evaluations materially change the performance or safety picture.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.