Thinking Machines Lab has made the weights of Inkling 2 Small downloadable, completing the transition from the model's November 5 hosted-only introduction. An explicit November 12 update to the company's launch article points to the new Hugging Face repository, where the checkpoint is labeled Apache 2.0. This is the material news for open-weight users: researchers and infrastructure teams can inspect and deploy actual model artifacts instead of depending solely on Tinker access. The original launch and the later weights-day event should remain distinct in a release history. The license on the downloadable files should also be read alongside any acceptable-use information from the publisher. Primary source ↗
Inkling 2 Small is a sparse mixture-of-experts multimodal transformer with 594 billion parameters in total and roughly 17 billion active per token. Its 48-layer decoder routes tokens to eight of 384 experts and supports text-plus-image input with text output. The model card describes up to one million tokens of context, whereas the earlier launch article discussed a 256K served context with compaction for long tasks. These statements should not be collapsed into a single guaranteed production window; memory budgets, attention implementation and the serving stack remain decisive.
The most practical qualifier is hardware. Thinking Machines distributes NVFP4 as the native checkpoint and MXFP8 as an alternative. The official minimum is 500GB of combined GPU VRAM for NVFP4 or 800GB for MXFP8. Examples include two B300 or four H200 GPUs for the smaller format. The company lists SGLang, vLLM, TokenSpeed and Unsloth integration paths. That is valuable openness, but it does not mean the model fits a consumer laptop or a single desktop GPU. Compatibility must be validated against the exact runtime build and repository revision.
Thinking Machines publishes agentic coding and reasoning comparisons for Inkling 2 Small, but these are the publisher's evaluation tables and should not be presented as independently replicated results for the downloadable formats. For the Open Model Weights evidence registry, immediate work is to pin the released checkpoint revision, enumerate the NVFP4 and MXFP8 file manifests, distinguish the two stated context conditions and verify the Apache license at artifact level. No automatic Leading designation or reconstructed performance benchmark follows from the weight release.
What this changes for the evidence layer
New material weight-release event: verify thinkingmachines/Inkling-2-Small revision SHA, artifact-level Apache-2.0 notice, weight-shard manifests, NVFP4/MXFP8 hardware caveats, served 256K versus card 1M context, original November 5 announcement versus November 12 weights publication. Do not infer independently verified scores or curation.
Thinking Machines Lab — Inkling 2 Small official model card and weight-release update
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗