Inkling 2 Small is downloadable now: open weights arrived a week after launch
Direct answer
On November 12, 2026, Thinking Machines Lab published the weights of Inkling 2 Small on Hugging Face under the Apache 2.0 license, a week after launching the model hosted-only on its Tinker platform. The model is a mixture of experts with 594 billion total parameters and 17 billion active per token, accepts text and image input, generates text and carries a 1-million-token context window. It scores 31 on the Artificial Analysis Intelligence Index. The download includes the main checkpoint in NVFP4 format and an MXFP8 variant, with vLLM support from day one.
What shipped on November 12
- The full weights on Hugging Face: Thinking Machines announced the download on X at 20:50 UTC on November 12, and Mira Murati shared it minutes later. The model card declares the Apache 2.0 license, which permits commercial use and redistribution.
- Two precision variants: the main checkpoint in NVFP4, across 25 safetensors files, and an MXFP8 variant. Both ship with a chat template and a transformers-ready configuration.
- The safe path to open weights: the lab promised at the November 5 launch that it would publish the weights after external safety testing and red-teaming. The November 12 update confirms the download; it describes no results from that testing.
- Serving support from day one: vLLM announced support with vision, mixed-precision mixture-of-experts layers and speculative decoding. API access runs through Together, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare.
What changes for your business
- The model downloads and runs on your infrastructure: open weights allow downloading, inspecting, modifying, fine-tuning and self-hosting the model. Control of data and latency stays in your cloud, and dependence on one supplier falls.
- API pricing is $0.30 input and $1.00 output: Together charges $0.30 per million input tokens and $1.00 per million output tokens. The hosted route stays available for teams that prefer not to run their own infrastructure.
- The score of 31 marks the model's level: on the Artificial Analysis Intelligence Index, Inkling 2 Small scores 31, against 13 for its predecessor Inkling Small. On Code Arena WebDev it rose from 1,408 to 1,647 points, first among US open models in Arena's snapshot.
- The Apache 2.0 license is the practical difference: it permits commercial use and includes an explicit patent grant. The model works for products, and also for research, under the same license.
What to do in your business this week
- Test the model on a repeated task: run your task on Tinker's hosted route and compare cost, latency and quality with your current model. Cost per task is the number that decides the switch.
- Price self-hosting: the weights are free under Apache 2.0, and the cost is the infrastructure to run them. Ask your team for the cost per million tokens in your cloud before choosing between hosted and self-hosted.
- Read the license before selling: Apache 2.0 permits commercial use, and the model carries a patent grant. Your legal adviser confirms the terms before the model enters a paid product.
- Keep this article next to the November 9 one: that article introduced the model with weights promised for the following week. The November 12 download is the promise kept, a week later.
Note: this article summarizes Thinking Machines Lab's November 12, 2026 update and coverage by AlphaSignal, Superpower Daily, CellCog and AI Primer. Scores come from the vendor and from Artificial Analysis. The lab published no results from its external safety testing.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- Inkling 2 Small
- The open model you can run on your own server: Aleph Alpha Kolibri (Apache 2.0)
- Mistral Large 4: the sovereign open model with 1 trillion parameters and a license worth reading
Frequently asked questions
What are a model's weights?
They are the learned values inside the model. Publishing the weights lets you download the model, inspect it, modify it, fine-tune it and run it on your own infrastructure.
When were Inkling 2 Small's weights published?
On November 12, 2026, a week after the model's November 5 launch on the Tinker platform. Thinking Machines Lab announced the download on Hugging Face under the Apache 2.0 license.
What license do Inkling 2 Small's weights carry?
Apache 2.0, which permits commercial use and redistribution and includes an explicit patent grant. The model works for products and for research under the same license.
How much does Inkling 2 Small cost by API?
Together charges $0.30 per million input tokens and $1.00 per million output tokens. The weights are free under Apache 2.0; the cost of self-hosting is the infrastructure to run them.
