Thinking Machines 发布 v2 小模型：68 小时预训练，NVFP4 全程训练

The weight-release plan matters more than the launch itself: teams can fine-tune on Tinker first, then self-host once weights open, avoiding lock-in. Practitioners evaluating small models for cost-sensitive deployments can use Tinker to test Inkling 2 Small against similarly sized open-weight models before committing. Its availability on day one shortens the usual wait between announcement and usable access.

官方账号Mira Murati (TML)@miramurati11 sourcesSource

Score-centering moving from a niche trick to an industry default signals how RL post-training recipes are converging, and the book chapter offers a derivation that clarifies why training-inference mismatch hurts policy optimisation. Practitioners implementing GRPO-style or StableReinforce-style pipelines can use it to stabilise advantage estimates. It pairs well with the book's existing chapters on reward modelling and policy gradients.

Nathan Lambert@natolambertSource

The pricing gap against Tinker suggests infra cost, not algorithms, is becoming a competitive lever for agentic research pipelines. Isolated algorithm, training and sandbox environments are built specifically to guard against reward hacking during self-improvement cycles, a design detail teams running autonomous R&D agents will study closely. Practitioners can benchmark the API against their current training stack before committing budget.

官方账号Salesforce AI Research@SFResearchSource