Thinking Machines Ships Inkling 2 Small, a 594B-Parameter Mixture-of-Experts Model
Thinking Machines has introduced Inkling 2 Small, a 594B-parameter Mixture-of-Experts model with 17B active parameters. It supports image inputs and variable effort. The model is initially available through Tinker with closed weights; full weights are expected under Apache 2.0 after safety testing.
Thinking Machines has announced Inkling 2 Small, the first model in its second generation. The model is a Mixture-of-Experts with 594B total parameters and 17B active parameters, accepting image inputs and supporting variable effort. It scores 31 on the Artificial Analysis Intelligence Index, far surpassing the company’s previous models. The company attributes the gains to more efficient pre-training and a large, unified RL run. Inkling 2 Small is available on Tinker with closed weights; Thinking Machines says it will release the full weights under Apache 2.0 after external safety testing and red-teaming. Users can try it on Tinker Playground, fine-tune and sample on Tinker, and download checkpoints when weights are released.
Agentic Coding and Autoresearch
Inkling 2 Small was trained with RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses and context-compaction strategies. The company increased quality of synthetic coding data and moved beyond unit-test rewards: many tasks included an agentic RL grader that evaluated how the model approached problems, not just pass/fail. Early checkpoints were tested on OpenCode under the name Fledge-Alpha.
The result, Thinking Machines says, is a model that follows instructions better, considers changes carefully, and verifies its work without being asked. On benchmarks, Inkling 2 Small posted 25.4% on FrontierCode v2.0 (second-highest reported), 72.1% on DeepSWE v1.1 (second-highest), 86.3% on SWEBench Pro v2 Hard (second-highest), 81.6% on Vibe Code Bench v1.1, 31.9% on PostTrainBench v1.2 (second-highest), and 5.7% on Terminal Bench Science v0.1.
On PMPP-Hard, an external GPU kernel-optimization benchmark with 69 tasks and a 10-hour budget, Inkling 2 Small solved a mean of 54.25 tasks with Pass@4 of 64 and AUC of 494.27 task-hours. That places it behind GPT-6 Astra (63), Kimi K3.1 Pro (60.5), and Claude Opus 5.5 (60), but ahead of GLM 5.3 (52.5) and Mistral Large 4 (50.25).
The model also was trained on long-horizon autoresearch tasks, including GPU kernel optimization and machine learning tasks like post-training a model autonomously. In a preliminary kernel optimization test on DeepSeek-V3.2 sparse attention forward+backward, Inkling 2 Small achieved a 6.02× speedup over the original code, compared to 18.80× for Claude Opus 5.5 and 10.26× for GPT-6 Astra.
In an end-to-end autoresearch demonstration, Inkling 2 Small autonomously trained Qwen3.5-4B on Tinker to forecast time series such as weather and Wikipedia page views. The run took 4.5 hours and $426 in Tinker spend. According to the task log, the base model had a skill of 0.191; after data curation, multiple SFT runs, an RL attempt, and evaluation cost optimization, the final SFT4 submission reached skill 0.289.
Vision and Visual Creation
Inkling 2 Small processes images with a 292M-parameter vision encoder based on Perception Encoder (Bolya et al., 2025). It uses 20×20-pixel patches with 2×2 spatial pooling and interleaves image patches with text tokens. Thinking Machines says the model shows strong performance in OCR, grounding, infographic interpretation, and visual reasoning.
On visual benchmarks, Inkling 2 Small achieved the highest reported score in several rows: MMMU Pro 81.6%, PerceptionBench 64.7%, Humanity’s Sixth Sense 33.3%, ScreenSpot Pro with Python 91.5%, BabyVision with Python 93.7%, and Chartography with Python 80.4%. On DesignArena Website, it scores 1,340, below Muse Wonder (1,366) and Muse Spark 1.3 (1,351) but above GPT-6 Astra (1,333). Pending weight release, it is set to be the top open-weights model on that leaderboard.
Visual capabilities extend to iterative creation. The company showed a city simulator built with Three.js procedurally, with no pre-made assets; a Python-based PC emulator that boots a real Linux kernel in the browser; a hand-drawn valley with origami cranes rendered in Blender; and a floor-plan walkthrough generated from a map. In a blind wine tasting test, Inkling 2 Small identified three wines from a single picture in 19 out of 20 trials, outperforming the company’s researchers.