Most labs pitch a new model as a better chatbot. Thinking Machines is pitching one as a better research intern. That’s the interesting part of Inkling 2 Small, and it’s a sharper positioning choice than the benchmark table suggests.
Mira Murati’s startup released the model on Nov. 5. It’s the first in the company’s second generation. It’s a mixture-of-experts design with 594 billion total parameters and 17 billion active per token, plus a 292-million-parameter vision encoder. You can sample and fine-tune it now on Tinker, the company’s training platform. Full weights arrive next week under Apache 2.0, after external safety testing and red-teaming.
The training numbers are specific. Pretraining covered 8.4 trillion tokens in 68 hours on Nvidia GB300 NVL72 racks. Then came a single nine-day reinforcement learning run on 2,048 GPUs across coding, agent and reasoning tasks. The target use is what the company calls “autoresearch”: post-training other models and optimizing GPU kernels. In one demo, the model fine-tuned Qwen3.5-4B for time-series forecasting on Tinker in about 4.5 hours. It spent $426 of a $500 budget and raised the skill score from 0.191 to 0.289.
That demo is useful, but it’s also a sales pitch for Tinker. Every autoresearch job runs on the company’s own platform, so the model doubles as a funnel. It’s a smart loop, and it’s different from what DeepSeek and Moonshot do, where the weights are the product and someone else’s infrastructure does the work.
On the company’s own benchmarks, Inkling 2 Small scored 72.1% on DeepSWE v1.1. That trails DeepSeek-V4.1-Flash at 77.0% and beats Kimi K3 at 68.5%. It led the listed open models on MMMU Pro (81.6%) and ScreenSpot Pro (83.8%). So it isn’t the best open coding model. It is a strong multimodal one, and the screen-understanding score fits an agent-focused pitch. These are self-reported numbers, so wait for independent evals once the weights are out.
The company is also candid about weak spots. It says the model still trails closed-weights models on alignment, and the chat experience is weaker, including overreacting to mental-health and safety queries. Anyone hoping to swap it into a consumer assistant should read that as a warning. Developers building internal tooling or training pipelines are the real audience.
The money matters here too. Business Insider reported the lab is raising $5 billion led by Accel at an expected $40 billion valuation, per SiliconANGLE. Thinking Machines hasn’t commented. The Information reported in September that Nvidia could supply about half of the round. The lab raised $2 billion at a $12 billion post-money valuation in July 2025, so the implied jump is more than threefold in about sixteen months.
That valuation needs a story beyond “another big open model.” Inkling 2 Small gives it one: own the post-training workflow, give it away at the weights level, and monetize the compute and tooling around it. Whether it works depends on the Apache 2.0 release next week and on whether outside teams can reproduce results like that $426 forecasting run. If they can, Tinker becomes a real alternative to stitching together open models and rented GPUs. If they can’t, this is a well-funded model with a good pitch.
