Thinking Machines released Inkling-Small on July 30, two weeks after the larger Inkling model. The new model has 276B total parameters and 12B active parameters, accepts text, image and audio inputs, and is available with open weights.
The company's release says Inkling-Small was trained after Inkling, using an improved data mix and training recipe. It also used on-policy distillation from Inkling and then two more weeks of agentic coding reinforcement learning.
“Inkling-Small’s combination of broad performance and efficiency will make it easy to experiment with and test in real applications.”
A smaller model, not a small deployment
The model card lists a 42-layer decoder-only Transformer with sparse MoE feed-forward layers. Each token routes to 6 of 256 experts, plus 2 shared experts. It supports a context window of up to 1M tokens and carries an Apache 2.0 license.
The practical limit is hardware. vLLM's recipe says the NVFP4 variant needs at least 180GB of aggregate VRAM, while BF16 needs at least 600GB. That is still far beyond a laptop, but it is far less forbidding than the original Inkling's multi-terabyte footprint reported by 36Kr/Xinzhiyuan.
The benchmark story has edges
Thinking Machines reports 80.2% on SWE-bench Verified, 64.7% on Terminal Bench 2.1, 31.6% on text-only HLE, and 40.1% on ARC-AGI-2. Each beats the larger Inkling in the same table.
Artificial Analysis gives the more cautious outside view: Inkling Small scores 40 on its Intelligence Index, one point below Inkling's 41, and no open-weight model at its size or smaller scores higher. It also notes weaker factual knowledge and some agentic-task gaps.
What to watch now
The pressure on other open-model labs is operational. Thinking Machines did not only publish a weight file; it published the model card, Hugging Face access and deployment paths through vLLM and SGLang. The next useful checks are whether teams can run it reliably at the 180GB/600GB tier, fine-tune it for their own codebases, and reproduce the coding and multimodal gains outside the official harness.
Sources: 36Kr/Xinzhiyuan, Thinking Machines release, CocoLoop, Inkling-Small model card, Artificial Analysis, Hugging Face and vLLM Recipes; official sources verify release date, parameters, license, context, training and benchmark scope, while third-party sources verify deployment requirements and independent scores.