Inkling 2 Small: A model that can see and refine what it builds
Thinking Machines Lab released Inkling 2 Small on November 5, 2026, opening the second generation of its Inkling family. The model uses a mixture-of-experts design with 594 billion total parameters, of which 17 billion activate for each token. It accepts image input, lets users set how hard it reasons on a task, and offers a context window of up to 1 million tokens, although Tinker's serverless tier lists 256K. The lab reports a score of 31 on the Artificial Analysis Intelligence Index, up from 13 for its predecessor, Inkling Small.
The network has 48 decoder layers and routes each token to 8 of 384 experts, paired with a 292 million parameter vision encoder. The lab says it pretrained the model natively in NVFP4 on NVIDIA GB300 NVL72 racks, then ran one reinforcement learning run over nine days on 2,048 GPUs. In one demonstration, the model autonomously trained a Qwen3.5-4B forecaster on Tinker to predict weather and Wikipedia page views, a job that took about 4.5 hours and $426 in spend, according to pasqualepillitteri.it.
Access is hosted for now. Inference costs $0.30 per million input tokens and $1.00 per million output tokens, and fine-tuning runs on Tinker. The weights stay closed until external safety testing and red-teaming are complete, and the lab expects to release them under Apache 2.0 next week. The model does not accept audio. The launch is also listed on Product Hunt.
Why it matters
For teams building on open models, the release changes the order of operations. Customers can fine-tune and serve the model through Tinker before deciding whether to self-host, and the lab has committed to a permissive license for the weights. Per-token prices are low, but output length drives real cost. Artificial Analysis reportedly measured about 150K output tokens per index task, roughly four times Inkling Small, according to theddari.
For investors, the launch shows a lab shipping a second-generation model less than four months after the first, and it makes license timing the item to verify. Self-hosting a 594B parameter model is not a small task: simple arithmetic puts a 4-bit copy near 297GB before runtime overhead, so the 17B active count does not shrink memory needs. Founders should check the license terms when the files arrive rather than assume openness from the company's stated stance.
Key points
- 1Inkling 2 Small has 594 billion total parameters and 17 billion active parameters per token.
- 2The model scored 31 on the Artificial Analysis Intelligence Index, compared with 13 for Inkling Small.
- 3Tinker charges $0.30 per million input tokens and $1.00 per million output tokens for the model.
- 4Model weights remain closed for now, with an Apache 2.0 release promised after external safety testing.
- 5
