Thinking Machines releases Inkling 2 Small weights under Apache 2.0
Thinking Machines Lab published Inkling 2 Small's full weights on Hugging Face under Apache 2.0.
· Last verified: 13 Nov 2026 (Thinking Machines Lab (Hugging Face), X)

Summary
- Thinking Machines Lab published Inkling 2 Small's full weights on Hugging Face under Apache 2.0.
- The model card lists 594 billion total parameters, 17 billion active, and a 1-million-token context window.
- Open weights let companies run and commercialize the multimodal model on their own servers and data.
The latest
The weights of Inkling 2 Small are now downloadable on Hugging Face under an Apache 2.0 license, allowing commercial deployment on private infrastructure. Thinking Machines Lab's Mira Murati said on X the release followed work by the company's team and its safety testing partners. The repository runs to 394 gigabytes of files.
Details
- The license: The model card lists Apache 2.0, one of the most permissive open-source licenses, which permits commercial use, modification and redistribution without a separate agreement with Thinking Machines Lab. That is the core difference from a hosted API, where usage runs through the provider's own terms and servers.
- The size: The card states 594 billion total parameters with 17 billion active per token. The Hugging Face page header shows a different figure, 383B params, drawn from Safetensors metadata, leaving two conflicting counts attached to the same release.
- The architecture: A 48-layer decoder transformer with a sparse mixture of experts, routing each token to 8 of 384 experts, combined with local and global attention layers. The design keeps active compute far below total parameter count.
- Inputs and context: The model accepts text and images and outputs text, with a context window of up to 1 million tokens. The card says image performance is best between 40 and 4096 pixels per dimension.
- The precision: Pre-training, mid-training and post-training ran in NVFP4, the recommended native precision. An alternative MXFP8 checkpoint gains 1 to 2 percentage points on evaluations, at a higher hardware cost.
- The hardware bar: NVFP4 requires at least 500 gigabytes of aggregate VRAM, for example 2 NVIDIA B300s or 4 H200s. MXFP8 requires at least 800 gigabytes, or 4 B300s or 8 H200s, placing self-hosting beyond most small teams.
- Runtime support: The card names SGLang, vLLM, TokenSpeed, Unsloth, Hugging Face Transformers and Docker Model Runner as supported serving frameworks.
- The scores: Benchmarks published by the company include Artificial Analysis v5.0 at 30.6, Vals Index 56.1%, SWEBench Pro v2 Hard 86.3%, BrowseComp 90.2% with compression, ARC-AGI-2 81.7%, HLE text-only 44.1%, OmniDocBench v1.6 94.6% and Global-MMLU-Lite 91.0%. All are self-reported.
- The safety review: Thinking Machines Lab said it assessed sycophancy, harmful manipulation, psychological harm patterns, chemical, biological, radiological, nuclear and cyber uplift, and loss-of-control risk, including refusal-suppressed variants. It concluded the model "did not present risk of material uplift beyond what's already available in the open-weight ecosystem."
- The limits: The card lists hallucination, occasionally imprecise instruction-following, degraded performance in long multi-turn conversations and a knowledge cutoff. It recommends human oversight for high-stakes use and warns against medical, legal or safety-critical deployment without additional tuning and verification.
Background
Thinking Machines Lab first launched Inkling 2 Small as a hosted model and committed to opening its weights. The release delivers on that commitment, moving the model from a provider-controlled interface to a downloadable artifact enterprises can run themselves.
Between the lines
The permissive license and the hardware floor pull in opposite directions. Apache 2.0 removes the legal barrier to commercial self-hosting, but a 500-gigabyte VRAM minimum and a 394-gigabyte download mean the practical audience is organizations already running multi-GPU clusters. The two conflicting parameter counts on the same page also leave outside evaluators without a settled figure to size the model against rivals.
What's next
Watch for independent benchmark runs testing the company's self-reported scores, clarification of the parameter discrepancy between the model card and the Safetensors header, and the first enterprise deployments on NVFP4 hardware.