Overview
Inkling 2 Small is the second-generation model from Thinking Machines Lab, released on November 5, 2026: a 594B total, 17B active MoE with a 292M vision encoder, image input support and variable effort, scoring 31 on Artificial Analysis Intelligence Index v5.0.1, 18 points above the previous Inkling Small. It is available for fine-tuning and inference on Tinker the day it shipped, with full weights scheduled for release under Apache 2.0 the following week.
Its center of gravity is long-horizon autonomous work. Evaluation runs allow trajectories up to 1M tokens, and training moved in four stages: H1 laid the base with 8.4 trillion tokens, H2 shifted to 3.8 trillion higher-quality, fresher tokens and pushed context to 512K, H3 added roughly 320 billion tokens of reasoning and agent trajectories to reach a 1M window, and H4 refined on about 34 billion tokens of the highest-quality trajectories, followed by 1,075 steps of joint reinforcement learning, more than 17 million rollouts, 2,048 GPUs and nine days.
Results are evenly spread. On coding and automated research, SWEBench Pro v2 Hard reaches 86.3%, DeepSWE v1.1 72.1%, Vibe Code Bench v1.1 81.6% and PostTrainBench v1.2 31.9%. On vision, MMMU Pro 81.6%, ScreenSpot Pro 83.8%, GDP.pdf 25.2% and PerceptionBench 64.7%. On knowledge work and finance, JobBench 70.5%, Vals Legal Index 73.9% and Vals Finance Index 61.5%. Long-context LongBench v2 comes in at 70.6%.
Variable effort is the practical design of this generation: training samples an effort level at random as a condition and adds a length penalty to the reward, so the model learns to trade speed against thoroughness on its own. Evaluation runs at the top effort setting of 0.99, and everyday calls can dial it down.
Key Features
- 594B total, 17B active MoE: A 292M vision encoder cuts images into 20 by 20 pixel patches, applies 2 by 2 spatial pooling and interleaves them with text tokens. Training runs natively in NVFP4.
- Trajectories up to 1M tokens: The RL training window is 256K, pretraining stage H3 used 1M, and evaluation allows 1M-token trajectories, with context compaction trained for even longer task chains.
- Variable effort: Effort level is sampled and conditioned during training with a length penalty in the reward, so one model moves between quick answers and deep reasoning.
- Fine-tune directly on Tinker: Sampling and fine-tuning are available on Tinker from launch day, and the model can even orchestrate its own fine-tuning, so nobody has to wait for the weights.
- Coding and automated research benchmarks: SWEBench Pro v2 Hard 86.3%, DeepSWE v1.1 72.1%, Vibe Code Bench v1.1 81.6%, PostTrainBench v1.2 31.9%. Long-horizon autonomous work is the design target.
- Vision and knowledge work: MMMU Pro 81.6%, ScreenSpot Pro 83.8%, GDP.pdf 25.2%, alongside JobBench 70.5%, Vals Legal Index 73.9% and Vals Finance Index 61.5%.
- Apache 2.0 weights planned: The path runs through internal safety testing before release, with full weights expected the following week under Apache 2.0 and checkpoints published alongside them.
Use Cases
- Agentic coding: repository-scale refactors, multi-turn tool use and test backfill, where a 1M trajectory means long tasks need not restart midway.
- Screen-operating agents: ScreenSpot Pro at 83.8% signals solid grounding for GUI automation.
- Visual document work: GDP.pdf 25.2% and Charxiv RQ 85.6% mean reports, drawings and charts can be read directly.
- Fine-tuning on your own data: pay per token on Tinker instead of standing up a training cluster.
- Long-horizon automated research: in one published sample the model trained Qwen3.5-4B for time series forecasting on its own, doubling the score in 4.5 hours for $426 in training and fine-tuning costs.
- Knowledge work pipelines: the legal and finance benchmarks show it can take on contract review and earnings analysis.
Pros
- 17B active against 594B total hits a practical ratio between capacity and per-step cost.
- The 1M trajectory length is the most substantive upgrade of this generation, so long tasks no longer need to be stitched together.
- Variable effort lets one model cover both fast answers and deep thinking without maintaining two models behind a router.
- Vision and coding are both strong, with ScreenSpot Pro and SWEBench Pro v2 Hard near the front.
- The Apache 2.0 plan is clean: self-host, commercial use and redistribution all follow once weights land.
- Tinker removes the need for your own cluster, so fine-tuning on your data starts at per-token pricing.
- Native NVFP4 training eases VRAM pressure when deploying on Blackwell-generation hardware.
Pricing
During launch week the only channel is Tinker, where you pay per token for sampling or fine-tune directly on the platform, with rates set separately for training and sampling by model tier; the Tinker documentation carries the exact numbers. Full weights are scheduled for Apache 2.0 release the following week, after which self-hosting turns the cost into compute. The published cost sample: an autoresearch time series task where training plus fine-tuning one model cost $426 over 4.5 hours, against Tinker budget caps of $500.
Summary
Inkling 2 Small treats long-horizon autonomous work as the main target: 594B total with only 17B active, trajectories up to 1M, and variable effort switching between quick and thorough on its own. Coding, vision and knowledge work benchmarks all land near the front, and the AA index score of 31 sits 18 points above the previous generation. Trying it on Tinker now and deciding on self-hosting once Apache 2.0 weights arrive next week is the low-friction path.
Version History
- Inkling 2 Small launches at AA index 31 with open weights planned (2026-11-05): Thinking Machines Lab released its second-generation Inkling 2 Small on November 5, 2026: a 594B total, 17B active MoE with a 292M vision encoder, image input, variable effort and native NVFP4 training, scoring 31 on Artificial Analysis Intelligence Index v5.0.1, 18 points above the previous generation. Evaluation allows 1M-token trajectories, and training ran in four stages totaling roughly 12.2 trillion pretraining tokens, closing with 1,075 steps of joint reinforcement learning, more than 17 million rollouts, 2,048 GPUs and nine days. Benchmarks include SWEBench Pro v2 Hard 86.3%, DeepSWE v1.1 72.1%, MMMU Pro 81.6%, ScreenSpot Pro 83.8%, LongBench v2 70.6% and JobBench 70.5%. Fine-tuning and inference are available on Tinker from launch, with full weights planned the following week under Apache 2.0; the company says that if completed it would be the highest-scoring open-weight model from outside China.