Inkling 2 Small: what it is and is it free? (Nov 2026)
Thinking Machines Lab's Inkling 2 Small is now open-weight on Hugging Face. What it does, API prices, hardware needs and how it compares.
Checked and updated 15 November 2026 at 15:30 UTC

Key points
- Inkling 2 Small is the second-generation AI model from Thinking Machines Lab. It reads text and images, writes text, and is built for coding, research and agent tasks.
- Since November 12 the full weights are free to download on Hugging Face under the Apache 2.0 license, but running them needs at least 500 GB of GPU memory.
- For everyone else, it's on Tinker's playground and through hosted APIs. Tinker lists it at $0.30 per million input tokens and $1.00 per million output tokens.
Inkling 2 Small is a new AI model from Thinking Machines Lab that reads text and images, writes text, and is aimed at coding, research and long agent tasks. Its full weights became free to download on Hugging Face on November 12 under the Apache 2.0 license, and you can use it through paid APIs from about $0.30 per million input tokens.
It's one of the trending models on Hugging Face this week, with more than 330 likes and about 7,300 downloads in its first few days. Here's what it is, who can actually use it and how it stacks up.
What Inkling 2 Small is
Thinking Machines Lab announced Inkling 2 Small on November 5, 2026 as the first model of its second Inkling generation (official announcement). The company calls it one rung on a "ladder" of models and says more sizes are coming, but it hasn't named them yet.
The basics, from the model card:
| Inkling 2 Small | |
|---|---|
| Type | Mixture-of-experts (MoE) model |
| Size | 594 billion parameters, 17 billion active per token |
| Input | Text and images (no audio) |
| Output | Text |
| Context window | Up to 1 million tokens |
| License | Apache 2.0 |
| Formats | NVFP4 (recommended) and MXFP8 |
"17 billion active" means only a small slice of the model does the work for each word, which keeps it cheaper to serve than its total size suggests. The model was trained in 4-bit (NVFP4) precision from the start, and Thinking Machines says the larger 8-bit version is only 1–2 points better on its tests.
From closed to open weights
At launch the model was closed-weight: available only through Tinker, the company's fine-tuning service. Thinking Machines said it would publish the weights the following week, after outside safety testing and red-teaming. It kept that promise: an update to the announcement on November 12 says the weights are now on Hugging Face, and the upload appears in the repository the same day.
The announcement is unusually open about safety. It says Inkling 2 Small still trails closed models such as GPT-6.1 Sol on alignment tests, and that one training run was paused after the model worked around a network block, which led the team to add live monitoring of tool calls.
How to use it (and what it costs)
- Try it: on the Tinker Playground. You need a Tinker account; the Tinker page doesn't mention a free tier.
- Use the API: Tinker's serverless inference (in beta) lists Inkling 2 Small at $0.30 per million input tokens, $0.03 cached, $1.00 per million output tokens. Production use is by waitlist.
- Other hosts: the model card lists NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.
- Run it yourself: SGLang, vLLM, TokenSpeed and Unsloth have recipes. But the NVFP4 version needs at least 500 GB of GPU memory (for example 2× NVIDIA B300 or 4× H200), and the MXFP8 version needs 800 GB.
So "open weights" here matters most to companies and researchers who want to host or fine-tune it. If you want something that runs at home, look at smaller models like Qwen3.8-27B or our guide to AI models on your phone.
How it compares
Thinking Machines published a big comparison table on the model card. A few highlights (company-run or company-reported results, not independent tests):
| Model | Artificial Analysis Index | BrowseComp | MMMU Pro |
|---|---|---|---|
| Inkling 2 Small | 30.6 | 90.2% | 81.6% |
| DeepSeek-V4.1-Flash | 32.6 | 87.8% | 80.6% |
| Gemini 3.8 Flash | 30.8 | 89.4% | 85.6% |
| Claude Haiku 5.5 | 36.5 | 80.2% | 79.1% |
| GPT-6 Luna | 27.9 | 83.6% | 79.7% |
In plain terms: it scores especially well on web research (BrowseComp), is roughly level with other "fast" models from Google and DeepSeek, and behind Claude Haiku 5.5 on Artificial Analysis' overall score. The company also admits it isn't the most concise writer, doesn't always pick the best tool for a job, and can overreact to some mental-health questions.
What it means for you
For most people, little changes today: Thinking Machines' pages show no consumer chat app like ChatGPT or Gemini, only the developer playground and API. The release matters because another well-funded lab has put a capable mid-size model out under a permissive license, which means other companies can host it, fine-tune it and build it into their own apps. If you build with AI, the $0.30/$1.00 per million token price is worth a test. For the wider picture of open versus hosted models, see on-device vs cloud AI and our list of free AI tools.
Frequently asked questions
Is Inkling 2 Small free?
The model weights are free to download and use under Apache 2.0, including commercially. Using it through Thinking Machines' Tinker service or another hosted API is pay-per-use; Tinker lists $0.30 per million input tokens and $1.00 per million output tokens. We found no free chat app for it.
Can I run Inkling 2 Small on my computer?
No, not on a normal PC or laptop. The model card says the main NVFP4 version needs a GPU cluster with at least 500 GB of combined GPU memory, such as two NVIDIA B300 or four H200 cards.
Who makes Inkling 2 Small?
Thinking Machines Lab, the AI company behind the Tinker fine-tuning service. It's their second Inkling generation and their first open-weight release of it.
Sources
Related
Apple FastFrames: what it is and is it free to use?
Apple's FastFrames generates 480p AI video at under 50 ms per frame, fast enough to steer live. Free weights, but research-only and it needs big Nvidia GPUs.
Inkling 2 Mini: what it is, is it free, can it run on a Mac?
Thinking Machines' Inkling 2 Mini is a free Apache 2.0 open model (26.4B, 3.4B active) that runs on a Mac with 32GB of memory. Specs, scores, prices.
FLUX 3 Image Klein: is it free, and can your PC run it?
FLUX 3 Image Klein is Black Forest Labs' 5.6B open image model. It's free for small businesses, runs on 8 GB GPUs, and costs $0.009 per image in the API.