Openness profile

Thinking Machines Inkling (Inkling 2 Small, 594B MoE; Inkling 975B and Inkling-Small from July 2026)

The first Apache-2.0 frontier-class weights release from a US lab, and it arrives on the lab's own safety schedule.

Model files · Inkling-2-Small huggingface.co

The six questions

Read the rubric in Chapter 13

Access

3 of 3 · Strong / exemplary

weights downloadable, ungated — but only after a staged period of closed, gated hosting on Tinker

Access in the book

Permission

3 of 3 · Strong / exemplary

Apache 2.0 on the released weights; the acceptable-use policy covers the hosted API, not the artifacts

Permission in the book

Deployment

2 of 3 · Yes, with caveats

day-0 ecosystem support including llama.cpp, but a 500 GB+ VRAM floor with NVFP4-first checkpoints

Deployment in the book

Transparency

2 of 3 · Yes, with caveats

phase-level recipe, data-category documentation, held-out bits-per-byte and 50+ published evals; corpus and training code withheld

Transparency in the book

Reproducibility

1 of 3 · Limited / partial

Limited to moderate.

Reproducibility in the book

Governance & Participation

2 of 3 · Real but bounded participation

The lab publishes its staged 'safe path to open weights' policy, documents external red-teaming, and ships day-0 community integrations — yet it alone decides when a model leaves the closed stage, and the staging means the flagship Inkling 2 sizes announced for release 'soon' are still pending.

Governance in the book

What is actually released

  • weights (Inkling 2 Small, 594B total / 17B active MoE, native NVFP4 plus MXFP8 checkpoint)
  • 292M Perception-Encoder-based vision encoder
  • day-0 support in transformers, vLLM, SGLang and llama.cpp
  • model card with a structured evaluation table against twelve competitors
  • phase-level training recipe (four pretraining phases plus a joint RL run) and a Training Data Documentation page; no training code, no corpus, no paper

Terms of use

Artifact license

Apache License 2.0 on the downloadable weights (NVFP4 and MXFP8 checkpoints); a separate Acceptable Use Policy governs only the hosted Tinker API

Read the license source huggingface.co

Artifact license: OSI-approved

OSI approval here describes the listed artifact license. It does not certify the full release under the Open Source AI Definition.

Separate access agreements

No additional download or access agreements are recorded for this assessment. Check the cited release terms.

Download gates and accompanying contracts are recorded separately from the artifact license’s approval status.

Assessment record

Assessment scope
Only the named artifacts and release versions listed in this profile; no score applies to other releases from this publisher.
Artifact kind
Trained model
Assessment recorded

Tier rationale

Permission/deployment gate: P<2 or D<3. A=3 P=3 D=2 T=2 R=1. G is displayed separately.

How the tier is assigned →

What the publisher claims

open weights from a frontier lab built around fine-tuning and adaptation (Tinker)

open weights, open science

Why it matters

The staged release — closed on Tinker, then downloadable after external red-teaming — is now the template for how a well-funded Western lab ships openness: real weights, real permissiveness, on a timeline the company controls.

Sources

Dataset snapshot 2026-11-15 · Source and history github.com . An assessment date records this profile; it does not establish that every source was rechecked that day. Scores are qualitative readings of the cited evidence; the tier follows the published rules and does not measure intelligence.

← Back to the field guideJoin the newsletter →