Kivorent News

Wednesday, November 18, 2026

Tech and AINews

Thinking Machines releases Inkling 2 Small's weights under an Apache 2.0 license

A week after launching Inkling 2 Small as a hosted-only model, Mira Murati's lab put its weights on Hugging Face. Anyone can now download, modify and self-host it, but only on hardware with roughly 500 GB of GPU memory.

Cortex, covering AI models, labs and agents
, 3 min read

594Btotal parameters; 17B active per token

What to know

  • Thinking Machines Lab published Inkling 2 Small's weights on Hugging Face on November 12 under an Apache 2.0 license.
  • The model launched on November 5 on the lab's Tinker platform with weights withheld pending external safety testing.
  • The main download is about 394 GB, and the model card calls for at least 500 GB of combined GPU memory.
Inkling 2 Small benchmark scores, Thinking Machines' evaluation
SWE-bench Pro v2 Hard86.3%
DeepSWE v1.172.1%
HLE (text only)44.1%
FrontierCode v2.025.4%

Source: OrcaRouter, November 13, 2026, citing Thinking Machines' published results.

Thinking Machines Lab, the AI startup led by former OpenAI technology chief Mira Murati, released the weights of its Inkling 2 Small model on Hugging Face on November 12 under the Apache 2.0 license. The license lets anyone download, modify and run the model, including for commercial products, without paying the lab.

The release follows through on a promise the lab made a week earlier. When Inkling 2 Small launched on November 5, it was available only through Thinking Machines' Tinker platform. According to Let's Data Science, the lab said at the time that it would publish the weights the following week, after external safety testing and red-teaming.

What was released

Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters, of which about 17 billion are active for each token, according to Superpower Daily and OrcaRouter. Each token is routed through 8 of 384 expert sub-networks. The model accepts text and images and has a context window of about 1 million tokens.

OrcaRouter, an API gateway company that examined the Hugging Face repository, reported that the first weights were committed at 19:53 UTC on November 12. The main repository, which stores most expert layers in Nvidia's compact NVFP4 format, is about 394 GB. A second version in the larger MXFP8 format is about 623 GB.

That size puts self-hosting out of reach of a single workstation. According to OrcaRouter and DEV Community, the model card calls for at least 500 GB of combined GPU memory for the NVFP4 version and 800 GB for MXFP8, which means a multi-GPU server such as two Nvidia B300s or four H200s. OrcaRouter also noted that the Apache 2.0 license is declared in the model card's metadata, but the repository had no separate license file when it checked, and serving it required development builds of the vLLM and SGLang inference engines.

How it performs

Thinking Machines' own evaluations put Inkling 2 Small at 86.3% on SWE-bench Pro v2 Hard and 72.1% on DeepSWE v1.1, two coding benchmarks, according to OrcaRouter. On DeepSWE and on FrontierCode v2.0, where it scored 25.4%, it trails DeepSeek-V4.1-Flash, Superpower Daily reported, while beating GLM-5.3-Flash, Muse Spark Open and Kimi K3. Independent firm Artificial Analysis gives it a score of 31 on its Intelligence Index, which OrcaRouter said ranks it fourth among 28 open-weight models. These are scores the lab chose to publish; see what a benchmark score does and does not tell you.

For people who do not want to run it themselves, Tinker charges $0.30 per million input tokens and $1.00 per million output tokens, and Together AI lists the same rates. Superpower Daily reported that Nvidia, Fireworks, Baseten, Nebius, Modal and Cloudflare also offer it through their APIs.

The safety testing question

The lab tied the weight release to outside safety testing, but none of the reports could find published results. OrcaRouter and Superpower Daily both describe the November 12 announcement as a short update linking to the Hugging Face repository, with no testing details. Once open weights are posted, the lab can no longer withdraw them or patch copies others have downloaded, which is why release-time testing draws attention.

Why it matters

Inkling 2 Small is a large, permissively licensed open-weight model from a US lab, and most of the rivals in its own comparison tables, such as DeepSeek, Kimi and GLM, come from Chinese labs. It gives companies that want to keep data on their own hardware, or fine-tune a model deeply, a permissively licensed American option, though the hardware bill means most will still rent it through a cloud API. For background on the trade-offs, see open-weight versus closed AI models. Thinking Machines is separately reported to be raising $5 billion at a $40 billion valuation.

Sources

AI modelsAI agents

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.