Thinking Machines Lab released Inkling on July 15, 2026, a 975 billion parameter mixture of experts model that marks the company’s first open-weights release. The move puts a fine-tuning specialist in direct competition with the open-weight labs whose models it has spent the past year helping developers customize. Inkling activates only 41 billion of its 975 billion parameters per token, the design choice that lets a model this large run at a cost closer to a system a fraction of its size.

The company built its reputation on Tinker, an API that lets developers fine-tune open-weight models without managing their own training infrastructure. Owning a flagship model changes that position. Thinking Machines is now both the maker of the tuning layer and the maker of a model meant to run through it, a vertical integration Mira Murati’s team had avoided until this release. Murati, the former OpenAI chief technology officer, founded Thinking Machines in 2025.

Inkling supports multimodal reasoning and a context window spanning one million tokens, roughly enough to hold a large codebase or several full novels in a single prompt. Thinking Machines also released a smaller preview version of the model, giving developers a way to test the architecture before committing compute to the full run through Tinker.

The release puts Thinking Machines inside a crowded open-weight tier: DeepSeek’s V3, Meta’s Llama family, and Moonshot AI’s Kimi K2 already compete there. All four use mixture of experts designs built to match closed frontier models on capability per dollar spent. Inkling’s 41 billion active parameters sit above DeepSeek V3’s 37 billion and below several of the larger Llama variants. That places Inkling as a mid-to-large entrant rather than the most compute-efficient option in the field.

Thinking Machines’ announcement does not include independent benchmark results comparing Inkling to DeepSeek V3, Kimi K2, or GPT-5 class systems on standard evaluations. Every open-weight model gets tested against the same public leaderboards within days of release. Inkling’s real test starts once that comparison exists.

The strategic logic behind the release is straightforward. A fine-tuning business needs base models developers actually want to tune, and depending entirely on Meta, DeepSeek, or Alibaba’s Qwen team for that supply leaves Tinker’s roadmap exposed to decisions made elsewhere. Owning a flagship open model gives Thinking Machines a captive product to sell customization around, plus a marketing surface the company lacked before July 15.

Teams evaluating open-weight base models for fine-tuning work should add Inkling to their benchmark set ahead of any Tinker contract renewal. Watch for third-party evaluations over the next two weeks to confirm whether the 41 billion active parameter figure holds up against DeepSeek V3 and Kimi K2 on real production workloads.

Per Thinking Machines’ announcement on July 15, 2026, introducing Inkling alongside its Tinker fine-tuning platform.