Tinker (Thinking Machines Lab)

Active, privately held.

Sources 1, 2, 3 · checked 2026-11-08

• Low-level training API (forward_backward, sampling) where your code defines the training loop and Tinker runs the GPUs
• LoRA training; recipes for SFT, RL environments, RLHF, and on-policy distillation
• Open-weight dense and MoE models from 4B to 1T parameters (for example Qwen3.5-4B, GPT-OSS-120B, Kimi-K2.6)
• API endpoint to download any saved checkpoint
• Serverless inference (beta)

Sources 1, 2, 3 · checked 2026-11-08

• Research and ML teams that want control of the training algorithm without running clusters
• Teams that need downloadable LoRA weights

Sources 1, 2, 3 · checked 2026-11-08

Per 1M tokens (as of October 19, 2026): Qwen3.5-4B (64K context) prefill $0.33, sample $1.005, train $0.737; GPT-OSS-120B (32K) prefill $0.33, sample $0.84, train $0.737; Kimi-K2.6 (32K) prefill $2.205, sample $5.49, train $4.84. Longer contexts cost more. Checkpoint storage $0.10/GB/month.

Sources 1, 2, 3 · checked 2026-11-08

None found / not verified. This is not a finding that Tinker (Thinking Machines Lab) lacks certifications; ask the vendor for documentation.

No certification found on the official site on 2026-11-08. Ask the vendor for security documentation.

Listed as stated by the vendor. Scope can differ by product or subsidiary; request current reports.

Sources 1, 2, 3 · checked 2026-11-08

• LoRA only; no full fine-tuning
• You write the training loop; no click-through managed job
• No published certifications found
• Some models are scheduled for retirement; check the model list

Sources 1, 2, 3 · checked 2026-11-08