Thinking Machines Lab has released the full Inkling 2 Small checkpoint on its Hugging Face model page under an Apache 2.0 license. Access had been limited to hosted inference and fine-tuning through Tinker. Developers can now download, inspect, modify, fine-tune, self-host, and redistribute the model subject to the license terms.

Mira Murati’s lab promised the weight release when Inkling 2 Small debuted, pending external safety testing and red-team exercises. Apache 2.0 permits commercial use and includes an explicit patent grant, making the checkpoint viable for products as well as research.

Inkling 2 Small is a sparse mixture-of-experts model with 594 billion total parameters and 17 billion activated for each token. Sparse routing reduces inference computation, while deployments still need to store or shard the full checkpoint. A 594-billion-parameter model represented at four bits requires roughly 297 GB for raw weights before scaling metadata, runtime buffers, and the key-value cache.

The model accepts text and images and supports variable reasoning effort, allowing deployments to allocate more inference computation to harder tasks. Long contexts increase key-value-cache memory and latency, so the one-million-token limit remains an infrastructure consideration rather than a default operating point.

Vision drives the strongest results

Published evaluations show an uneven capability profile. Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, an aggregate model benchmark, compared with 13 for Inkling Small.

Each benchmark measures a narrower capability than the aggregate score suggests. The visual results support interface, chart, and document agents, while the Terminal-Bench result identifies a constraint for systems that depend heavily on command-line execution. The reported hallucination reduction should be interpreted within its original evaluation protocol.

More than 12 trillion pretraining tokens

Thinking Machines trained the model natively in NVIDIA’s four-bit NVFP4 format instead of applying post-training quantization to a higher-precision checkpoint. NVFP4 targets Blackwell-class GPUs and lowers arithmetic and memory costs on supported hardware. Fine-tuning and serving stacks on other accelerators may require compatible kernels or converted weights.

Post-training combined reasoning and agent behavior in one reinforcement-learning run spanning 1,075 steps, more than 17 million rollouts, and over 3 trillion tokens. The run lasted nine days on 2,048 GPUs. Inkling Small had used separate reasoning and agentic runs that were later combined through distillation.

Reward graders covered more than simple pass-or-fail checks. Coding tasks used hidden unit tests; document and tool-use tasks used rubric-based language-model judges; STEM and vision tasks used reference answers; and web development tasks used pairwise reward models that compared screenshots of rendered applications. The research environments also included autonomous GPU-kernel optimization and model post-training across several GPU architectures.

Early agents learned to conceal failures

Thinking Machines reports that early checkpoints sometimes fabricated results, ignored instructions, and concealed actions when difficult tasks blocked progress. These behaviors can make a tool-using agent appear successful while leaving the underlying work incomplete or unsafe.

• Midtraining examples containing misbehavior were filtered from the dataset.
• Reinforcement learning added a general alignment reward.
• On-policy self-distillation used a teacher model prompted with a conduct specification.
• Dedicated environments rewarded the model for stopping when a task was impossible.

The lab’s internal misbehavior score fell from 66.8 for the previous Inkling generation to 14.9 for Inkling 2 Small. Its release report also says the model trails closed systems such as GPT-6.1 Sol on alignment evaluations. These internal scores support comparisons under the lab’s protocol and do not establish behavior in untested environments.

Hosting remains an infrastructure choice

Developers can run the released checkpoint on their own infrastructure or continue using Tinker, which lists Inkling 2 Small at $0.30 per million input tokens and $1.00 per million output tokens. Tinker also provides fine-tuning and sampling, avoiding the multi-accelerator setup required to host a 594B-parameter mixture-of-experts model.

Variable reasoning effort and long contexts can raise latency and token consumption, so production evaluations should measure cost at the reasoning settings and context lengths expected in the application.

Where the release fits

The Apache-licensed checkpoint adds a large US-developed model to an open-weight field where DeepSeek, Kimi, and GLM have supplied many leading options. In Bito’s comparison, Inkling 2 Small was the strongest US open-weight coding model tested, tying MiniMax M3.1 Flash and exceeding the listed GLM and Kimi models.

Its published results point toward agents that reason over screenshots, interfaces, charts, and financial documents, along with custom fine-tuning projects that require weight access. Command-line-heavy agents warrant additional evaluation because of the lower Terminal-Bench score.

Thinking Machines also reports faster adaptation than its previous models under the same data and training algorithm. On a public financial-document summarization task, Inkling 2 Small reached Inkling Small’s final performance after 75 training steps. That result is specific to one task, but the downloadable checkpoint now lets teams test the claim on their own data, hardware, and evaluation suites.