What is Inkling?
Inkling is the first model from Mira Murati's Thinking Machines Lab, released on 15 July 2026 with full weights under Apache 2.0. It is a mixture-of-experts model with 975 billion parameters, 41 billion active per token, pretrained on 45 trillion tokens of text, images, audio and video. It reads text, images and audio in one network and writes text; it supports up to 1 million tokens of context.
The lab sells it mainly as a base for customisation through Tinker, its fine-tuning service, which offers it at 64K and 256K context. A smaller Inkling-Small (276 billion parameters, 12 billion active) followed in late July, and Inkling 2 Small reached Tinker on 5 November 2026 with weights promised after outside safety testing.
How much does Inkling cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| OpenRouter | $1.00 / $4.05 per million tokens | Input / output, via Together |
Free weights under Apache 2.0. OpenRouter (Together): $1.00 per million input tokens, $4.05 per million output, 524,288-token context. Tinker prices per its docs.
Is Inkling worth it?
Inkling is a sound open base model whose own makers say it is not the strongest available, open or closed, and its vision scores bear that out. On the lab's own table it scores 73.5% on MMMU Pro against 79.0% for Moonshot's Kimi K2.6 and 82.0% for Gemini 3.1 Pro, and 78.1% on CharXiv reasoning against 80.4% for K2.6. Kimi K2.6 is itself a generation old now.
Audio is where it is unusual. Thinking Machines calls it among the strongest open-weights audio models, and Inkling scores 91.4% on VoiceBench and 77.2% on MMAU, against 94.3% and 82.5% for Gemini 3.1 Pro. If you want one open model that can read a screenshot and a voice note in the same request, the list is short.
The Apache 2.0 licence is the other draw. It carries no revenue caps, regional carve-outs or output restrictions, which several rival open licences do. On OpenRouter it costs $1.00 per million input tokens and $4.05 per million output, cheap for its size, and it drew 1,792,792 Hugging Face downloads in the last month.
Self-hosting is a data-centre job. The BF16 weights need about 2TB of GPU memory, eight B300s or 16 H200s, and even the NVFP4 build needs 600GB, four B300s or eight H200s. Inkling-Small claims comparable performance at 276 billion parameters and fits on one B300 in NVFP4, which makes it the more practical download for most teams.
Where it earns its place
A call-centre analytics vendor
Answer questions over recorded calls and the screenshots agents attach
Send WAV audio (16kHz, ideally under 20 minutes) and images in one request through OpenRouter. At $1.00 per million input tokens, the cost is mostly input; Apache 2.0 lets the vendor resell the results.
An ML team at a legal-tech company
Fine-tune an open model on its own document workflows and keep the weights
Train on Tinker at 64K or 256K context, then export and self-host. Running the full model needs a multi-GPU Blackwell or H200 node, so many teams will fine-tune Inkling-Small instead.
A research group comparing open multimodal models
Study an encoder-free model that handles audio and vision natively
Full weights, day-one support in Transformers, SGLang, vLLM and llama.cpp, and a published architecture make it easy to inspect. The lab warns that video performance has not been evaluated.
Before you commit
- Setup
- Weights on Hugging Face in BF16 and NVFP4; recipes for SGLang, vLLM (0.26 or later), TokenSpeed, Unsloth and Transformers. Tinker Playground for a quick try.
- Integrations
- Transformers, SGLang, vLLM, llama.cpp, Unsloth GGUF builds, Tinker API, OpenRouter (Together), Hugging Face Inference Providers.
- Data and privacy
- Apache 2.0; self-hosted, nothing leaves your servers. Training data includes public web content, third-party data and synthetic data. OpenRouter's listing is not available through EU in-region routing.
- Cost at scale
- On OpenRouter, 1 billion input tokens and 100 million output tokens a month cost about $1,405. Self-hosted, the NVFP4 build needs four B300s or eight H200s.
- Limits
- OpenRouter context 524,288 tokens, Tinker 64K or 256K, full 1M only self-hosted; images ideally 40 to 4,096 pixels a side; audio ideally under 20 minutes; text output only.
What’s good
- Apache 2.0 with no usage restrictions
- Text, image and audio input in one model; 1M-token context
- $1.00 / $4.05 per million tokens on OpenRouter
- Fine-tuning service from the maker through Tinker
What’s not
- 73.5% on MMMU Pro, behind Kimi K2.6's 79.0%
- About 2TB of GPU memory in BF16, 600GB in NVFP4
- Video input exists but has not been evaluated
- Smaller Inkling-Small claims comparable results at a quarter of the size
Best for
Teams that want a permissively licensed open model reading text, images and audio, and plan to fine-tune it rather than use it as is.
Inkling alternatives
Inkling: common questions
- Is Inkling worth it?
- Inkling is the large open model you pick for its licence and its ears, not for its eyes. We rate it 3.8 out of 5.
- How much does Inkling cost?
- Free weights under Apache 2.0. OpenRouter (Together): $1.00 per million input tokens, $4.05 per million output, 524,288-token context. Tinker prices per its docs.
- What are the downsides of Inkling?
- 73.5% on MMMU Pro, behind Kimi K2.6's 79.0%; about 2TB of GPU memory in BF16, 600GB in NVFP4; video input exists but has not been evaluated; smaller Inkling-Small claims comparable results at a quarter of the size.
- Who is Inkling best for?
- Teams that want a permissively licensed open model reading text, images and audio, and plan to fine-tune it rather than use it as is.
- How much does Inkling cost for a team or in production?
- On OpenRouter, 1 billion input tokens and 100 million output tokens a month cost about $1,405. Self-hosted, the NVFP4 build needs four B300s or eight H200s.
- Is Inkling safe for company data?
- Apache 2.0; self-hosted, nothing leaves your servers. Training data includes public web content, third-party data and synthetic data. OpenRouter's listing is not available through EU in-region routing.
- What are the best Inkling alternatives?
- The closest alternatives we have reviewed are Kimi K3, Nemotron 3.5 Super VL, Gemma 4 and Qwen3.8-27B.
Answers checked Nov 7, 2026.