Tinker

Thinking Machines Lab‘s fine-tuning platform, and the reason Inkling‘s weights are free. It hosts Inkling for fine-tuning and serves it via an OpenAI-compatible chat-completions API (bearer-token auth, text and image input), with context options at 64K and 256K against the model’s full 1M inkling-announcement. Priced at a limited-time 50% discount at launch; rates live in the lab’s docs rather than the announcement.

Why it’s on the map

Most of this spoke’s providers monetize inference — you rent tokens (llm-api-pricing). Tinker monetizes adaptation: the lab gives away an Apache-2.0 base it openly says isn’t the strongest, then charges to customize it. That inverts the usual open-weight logic, where free weights are developer-mindshare seeding for a paid frontier tier (google‘s gemma-4/gemini split). Here there is no frontier tier to seed — the platform is the business.

Compare mistral-ai‘s Forge (custom training inside its sovereign platform): two labs converging on training-as-a-service over their own open weights, one arriving from a sovereignty play and one from having no other product at all.

Related

thinking-machines-lab · inkling · inkling-announcement · llm-api-pricing · open-weight-models · mistral-platform