Thinking Machines launches Inkling 2 Small, with open weights pending safety reviews
For the CTO of a bank weighing on-premises models
Thinking Machines Lab, the AI startup led by former OpenAI chief technology officer Mira Murati, introduced Inkling 2 Small on Nov. 5 and made it available for inference and customization through its Tinker platform, VentureBeat reported.1
The model combines 594 billion total parameters with a sparse architecture that activates 17 billion for each token, according to VentureBeat's report; the source does not say why the company built it at this size.1 VentureBeat reported that training included software repositories, terminal environments, document production, research, visual tasks and simulated interactions with users.1
On Surge AI's EnterpriseBench-CoreCraft benchmark, which measures agent performance across a fictional PC parts supplier, the model scored 71.8%, third behind Anthropic's Claude Fable 5.1 (77.4%) and Claude Opus 5.5 (74.4%).1
Per the company's announcement as relayed by VentureBeat, Artificial Analysis gave it an Intelligence Index score of 31, against 26 for Mistral Large 4; that would make it the top open model not from China once the weights are released.1
Thinking Machines plans to release the weights under Apache 2.0, but they remain closed at launch pending external safety evaluations, and the company targets release for the week after the announcement.1
The model accepts images and text but not audio, and VentureBeat lists Tinker hosted pricing at $0.30 input and $1.00 output per million tokens.1
The benchmark that produced the 71.8% score simulates a PC parts supplier, not a bank, so how the model performs on lending, KYC or servicing workflows is unknown and needs your own testing. Open weights under Apache 2.0 could allow self-hosting, but the hardware needs for a 594-billion-parameter model are a separate question for infrastructure teams.
Watch for the weight release, which Thinking Machines targets for the week after the Nov. 5 announcement, once external safety evaluations finish.1
Our take. A non-Chinese open model now sits a few points behind Claude on agent tests, and Apache 2.0 weights would let you run it in-house. 'Small' means 594 billion parameters, so price the servers first.
Action. Infrastructure engineering: Size GPU memory and server cost to host a 594-billion-parameter model before approving an on-premises Inkling 2 Small pilot.
From our records
- Claude (Anthropic)vendor · banks using it (21) Goldman, JPMorgan, BankUnited · alternatives Microsoft Copilot, ChatGPT Enterprise, Gemini Enterprise
- ChatGPT Enterprise (OpenAI)vendor · banks using it (14) Morgan Stanley, BNY, Customers Bank · alternatives Microsoft Copilot, Claude, Gemini Enterprise