Introducing Inkling 2 Small

Open original

Thinking Machines has released Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active parameters that accepts image input. It's available now through their Tinker platform, with full weights promised under Apache 2.0 after external safety testing.

  • The model scores 31 on the Artificial Analysis Intelligence Index, up from the company's previous generation, with gains credited to more efficient pre-training and a large unified RL run.
  • Coding results are mixed rather than dominant: it places second on several benchmarks like FrontierCode v2.0 (25.4%) and DeepSWE (72.1%), behind DeepSeek-V4.1-Flash on most rows.
  • On a self-run autoresearch test, the model autonomously fine-tuned Qwen3.5-4B over 4.5 hours for $426 in Tinker spend and more than doubled its forecasting score.

Inkling 2 Small is the second generation from Thinking Machines, a Mixture-of-Experts model with 594B total parameters, of which 17B are active per token. It accepts image inputs, supports variable reasoning effort, and scores 31 on the Artificial Analysis Intelligence Index, well above the company's first-generation models. The gains are attributed to more efficient pre-training and one large, unified RL run. For now it's available only through Tinker, with closed weights, including a playground for trying it and options for fine-tuning and sampling. The company says full weights under an Apache 2.0 license are planned for the following week, pending external safety testing and red-teaming. The most concrete claims are about coding and agentic work. The model was trained with RL on hundreds of thousands of long-horizon coding tasks across several agent harnesses, and it was graded not only on whether code passed unit tests but on how it approached the problem. In the comparison tables, it lands second-highest on several coding benchmarks: 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1, and 86.3% on SWEBench Pro v2 (Hard), trailing DeepSeek-V4.1-Flash and Kimi K3 respectively on those rows. It was also trained on autoresearch tasks like GPU kernel optimization. On the kernel benchmark, it solved 54.25 of 69 tasks on average within a 10-hour budget, close to but not at the top of the pack. One demo stands out: the model autonomously post-trained Qwen3.5-4B on Tinker to forecast time series, more than doubling its forecasting score in 4.5 hours for $426 of compute. On vision, the model uses a 292M-parameter encoder and reports top scores on several benchmarks, including 81.6% on MMMU Pro and 64.7% on PerceptionBench, along with leading numbers on ScreenSpot Pro and chart-reading tests. The company also cites a fun internal test where it identified three wines from one photo in 19 of 20 trials, beating its own researchers. Most of these results are self-reported and compare against a handful of rivals chosen by the company, so independent evaluation will matter more than the headline wins. The weight release next week is the real test: that's when outside developers can check whether the coding and vision numbers hold up in their own workloads.

Ask about this article