Inkling 2 Small now available on AI Gateway

Vercel's AI Gateway, a single API that forwards requests to many providers' models and tracks their usage and cost, now serves Inkling 2 Small under the model name thinkingmachines/inkling-2-small. For a developer already using the gateway, trying the model means changing that one string: the same AI SDK call works, so does an OpenAI-compatible Chat Completions client pointed at the gateway, and coding agents such as Claude Code, Codex and Cursor can select it after running Vercel's setup command. Vercel says the gateway charges the underlying provider's price with no markup and no platform fee, including when a team brings its own provider key. The changelog's capability description (long coding tasks, tool use, reading charts and screenshots) is Vercel's restatement of the lab's claims, not a separate test.

Inkling 2 Small is the second-generation small model from Thinking Machines, the lab behind the Tinker fine-tuning platform. Its release post describes a mixture-of-experts model, meaning each token passes through only part of the network: 594 billion parameters in total, 17 billion active, taking text and images as input and producing text. The lab says it wanted the model to be "a strong base model for customization." It launched on November 5 on Tinker with closed weights and a stated plan to publish them under Apache 2.0, and a November 12 update to the same post says the weights are now downloadable from Hugging Face and that hosted APIs are available at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. That is the practical background to the gateway listing. Vercel is not hosting a model only Thinking Machines can serve; its model page lists Together AI, Modal and Nebius as providers, so the gateway sits in front of several hosts running the same open weights, with one key, one bill and the gateway's retry and failover settings across them.

Independent measurements published at launch suggest where the model is worth trying and what it costs to run. Artificial Analysis scored it 31 on its Intelligence Index, level with Gemini 3.8 Flash (high) and 18 points above the first Inkling Small, with its largest gains on agentic tasks. The same evaluation found it used about 150,000 output tokens per index task, roughly four times its predecessor, so it cost $1.18 per task against $0.20. Low per-token prices still kept that below Gemini 3.8 Flash (high) at $7.77. Latent Space's roundup of launch reactions relays Vals AI's view that it is strong in finance and legal work but has ground to make up on technical agentic tasks, and the lab itself says the model is not the most succinct communicator and does not always choose the best tool for the job, especially in codebases with many skills. Those per-task costs were measured before the gateway listing, and gateway prices follow whichever host serves the request, so the token count is the number to budget around: a cheap rate per token can still add up on long agent runs.