Model components / Fine-tuning code

Tinker

Managed training API for researchers that lets teams control the training loop while Thinking Machines handles the infrastructure. It exposes four core primitives (forward_backward, optim_step, sample, and save_state) and supports frontier open-weight models including Qwen, GPT-OSS, Llama, DeepSeek, and Nemotron. Scale signal: Thinking Machines announced a multi-year, gigawatt-scale strategic partnership with NVIDIA and said NVIDIA made a significant investment to support the company’s long-term growth.

Tinker, Thinking Machines Lab's managed LoRA fine-tuning/training API (announced Oct 1 2025; now paid, usage-priced per-million-tokens + $0.10/GB-mo storage). Fine-tunes open-weight models 1B-550B (Qwen, Llama, DeepSeek, GPT-OSS, Moonshot, Nemotron incl. Qwen3-235B-A22B). Only the tinker-cookbook (Apache-2.0) is open; the training engine is a hosted proprietary service. Confirmed live June 2026.

Openness

1 high confidence

1.0

license: Proprietary
source: closed
engine: closed(hosted distributed-training service on TML infra)
cookbook: open(Apache-2.0, thinking-machines-lab/tinker-cookbook)
api: public-but-proprietary(forward_backward/optim_step/sample/save_state)
method: LoRA(+RL via sample)
access: paid+account-gated

Core training service is a proprietary hosted API; only the example cookbook is OSI-licensed, so the product is closed. Not open_core since no self-hostable engine exists.

https://thinkingmachines.ai/tinker/ recorded 2026-06-04
paid managed LoRA training API, usage pricing, supported models 1B-550B
https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/LICENSE recorded 2026-06-04
Apache-2.0 license for the (only) open component, the cookbook

Adoption

2 low confidence

2.0

Early named research adopters (Princeton Goedel, Stanford Rotskoff, Berkeley SkyRL, Redwood Research) and moved from private beta to paid GA; no disclosed user/usage count. Reported research traction => conservative level 2.

https://thinkingmachines.ai/tinker/ recorded 2026-06-04
paid GA service with named research adopters; no published user count

Capability

4 medium confidence

4.0

Strong fine-tuning capability (frontier-scale model breadth and managed distributed training) but LoRA-centric (not full FT / full RL stack), so 4 not 5. Feature-matrix basis; no MLPerf-Training submission.

https://thinkingmachines.ai/tinker/ recorded 2026-06-04
LoRA training across Qwen/Llama/DeepSeek/GPT-OSS/Moonshot/Nemotron, 1B-550B, MoE support

Unchanged since 2026-07-30 (last edited, not re-checked)