Thinking Machines Opens Inkling 2 Small Weights: 594B Parameters Under Apache 2.0

What happened

The model card for Thinking Machines Lab's Inkling 2 Small now carries the full checkpoint on Hugging Face, letting developers download, inspect, modify, fine-tune, self-host and redistribute it under the Apache 2.0 license. The model takes text and image inputs, generates text, and supports variable reasoning effort.

Before the weights went up, Inkling 2 Small was available only as hosted inference and fine-tuning through the lab's Tinker platform. The lab promised the weight release when the model debuted, pending external safety testing and red-team exercises; Tinker lists hosted pricing at $0.30 per million input tokens and $1.00 per million output tokens. Self-hosting runs in parallel, with local inference through frameworks such as SGLang and vLLM.

The model ships in two checkpoint formats: the main NVFP4 checkpoint needs at least 500 GB of aggregated VRAM, for example 2 B300 or 4 H200 GPUs, while the alternative MXFP8 checkpoint needs at least 800 GB. The lab says the model was pretrained on more than 12 trillion tokens, and its post-training ran a single reinforcement-learning run of 1,075 steps with more than 17 million rollouts.

Key facts

1. Architecture: A sparse mixture-of-experts model with 594B total parameters and 17B activated per token, a 48-layer decoder-only transformer routing each token to 8 of 384 experts; context window up to 1M tokens
2. Aggregate scores: 31 on the Artificial Analysis Intelligence Index, against 13 for the previous Inkling Small; hallucination rate 27%, down from 61% for Inkling Small
3. Single benchmarks: FrontierCode 2.0 at 25.4%, DeepSWE at 72.1%, ScreenSpot Pro at 83.8%, PerceptionBench at 64.7% and JobBench at 70.5%, all figures published in the lab's model card
4. Coding comparison: 76% in Bito's agentic coding benchmark, tied with MiniMax M3.1 Flash and the highest among US open-weight models tested; Terminal-Bench 4.0 at only 13%
5. Training scale: Pretrained on more than 12 trillion tokens, trained natively in NVFP4; post-training was a single reinforcement-learning run of 1,075 steps with more than 17 million rollouts
6. Safety assessment: Safety evaluations and external red-teaming were completed before release, and the lab concluded the model presents no risk of material uplift beyond the existing open-weight ecosystem; alignment evaluations still trail closed systems such as GPT-6.1 Sol