Thinking Machines released Inkling 2 Small, a 594B open-weight model

Thinking Machines Lab released Inkling 2 Small this week. The model is a Mixture-of-Experts with 594 billion total parameters and 17 billion active. The first release is on the lab's Tinker platform. Thinking Machines says the full weights will be published under an Apache 2.0 licence next week, after external safety testing.

Latent Space covered the release on the day of launch, in its daily recap.

What the model is

Inkling 2 Small reads text and images, and writes text. The lab says the vision encoder has 292 million parameters, based on Perception Encoder.

Thinking Machines says the main pretraining phase used 8.4 trillion tokens. It ran for 68 hours on NVIDIA GB300 NVL72 racks. Two further phases added 3.8 trillion and about 320 billion tokens. A final phase added 34 billion tokens of high-quality agent trajectories.

The reinforcement learning run was one joint run. It took 1,075 steps over 9 days on 2,048 GPUs. The lab says it used more than 17 million rollouts and over 4 × 10²³ FLOPs. Thinking Machines ran the whole pipeline in NVFP4, a 4-bit floating-point format from the Blackwell architecture.

The benchmark scores

Every number below is Thinking Machines' own, reported on the release page.

The Artificial Analysis Intelligence Index gives the model 31 points. On FrontierCode v2.0 it scores 25.4 percent. DeepSeek V4.1 Flash scores 28.5 percent and Kimi K3 scores 17.3 percent. On SWEBench Pro v2 Hard it scores 86.3 percent. Kimi K3 scores 88.2 percent. On DeepSWE v1.1 it scores 72.1 percent, behind DeepSeek V4.1 Flash at 77.0 percent.

The model wins on several vision and work tests. PerceptionBench is 64.7 percent, the highest in the compared set. ScreenSpot Pro is 83.8 and 91.5 percent on the two variants, both the best. JobBench is 70.5 percent, the highest. The Vals Finance Index is 61.5 percent, the highest. Vals CyberBench is 72.7 percent, the highest.

What the lab says is still weak

Thinking Machines says the model writes long answers. It says the model picks the wrong tool in some codebases with many skills. The model does not support audio input. The lab says chat quality is lower. Formatting is worse and the model overreacts on mental health and safety prompts. On alignment tests, the lab says Inkling 2 Small still trails closed models such as GPT-6.1 Sol.

The team shows one agent run as an example. It let the model train Qwen3.5-4B on time-series forecasting on its own. Thinking Machines says the run took 4.5 hours. The cost was $426 on the Tinker platform. The run more than doubled the baseline forecasting score.

A team running open-weight models at work gets a new candidate to test next week. Thinking Machines gives its own numbers to compare against. A fair test is one the team runs itself, on its own task, after the weights land.

Source

• Introducing Inkling 2 Small, Thinking Machines Lab.
• AINews: Inkling 2 Small, Latent Space.