Thinking Machines Releases Inkling 2 Small's Open Weights
Thinking Machines has released Inkling 2 Small's 594B-parameter weights under Apache 2.0. What it is, who can run it and what it means for African developers.

Thinking Machines Lab, the AI company led by former OpenAI chief technology officer Mira Murati, has made the full weights of its Inkling 2 Small model available to download. The company launched the model on 5 November 2026 as a hosted-only service, promised the weights "next week", and on 12 November updated its announcement to say they are now on Hugging Face under the Apache 2.0 licence, which allows commercial use.
Inkling 2 Small has 594 billion parameters, so "small" is relative: you cannot run it on a laptop. This article explains what was released, what open weights actually means, who can realistically use a model this size, and what it could mean for developers and startups in Nigeria and across Africa.
What Happened
The launch on 5 November
Thinking Machines announced Inkling 2 Small on 5 November 2026 as the second generation of its Inkling models. At launch, the weights were closed. People could use the model through Tinker, the company's own platform for fine-tuning and running models, but could not download it.
In its announcement, Thinking Machines said it would release the full weights under Apache 2.0 "next week", after external safety testing and red-teaming.
The weights arrive on 12 November
The company kept that promise. An update dated 12 November 2026 on the announcement page says the weights are now downloadable on Hugging Face. The official model card lists two versions:
- Inkling-2-Small, the main checkpoint, stored in a compact number format called NVFP4.
- Inkling-2-Small-MXFP8, a larger version in a different format, MXFP8.
Thinking Machines' own social media account said the weights were "now available for download on HuggingFace under an Apache 2.0 license", according to HuggingNews, which also reported that Murati described the release as happening "as promised".
The funding backdrop
The release comes as the company reportedly raises more money. Business Insider reported on 6 November, citing unnamed sources, that Thinking Machines was nearing a funding round of about US$5 billion at a US$40 billion valuation before the new money, led by the investment firm Accel. Thinking Machines declined to comment, according to that report, and we found no confirmation from the company. Treat this as reported, not confirmed.
The Details
What kind of model it is
According to the model card, Inkling 2 Small is a "Mixture-of-Experts" model. In plain terms:
- It has 594 billion parameters in total. Parameters are the numbers a model learns during training; more usually means more knowledge and skill, but also more memory needed.
- Only 17 billion of those are active for each word-piece (token) it processes. The model is split into 384 "experts", and for each token it routes the work to 8 of them. That makes it faster to run than a model of the same size that uses every parameter every time, but you still need enough memory to hold all 594 billion.
- It accepts text and images as input and produces text only. The first Inkling generation's audio support has been dropped, according to reports.
- It can handle up to 1 million tokens of context, which is roughly several long books' worth of text at once.
If parameters and tokens are new to you, our explainer on what an LLM is covers the basics.
The hardware you need
This is the part that matters most for anyone thinking of running it. The model card states:
| Version | Minimum combined GPU memory | Example setup (per model card) |
|---|---|---|
| NVFP4 (recommended) | At least 500 GB | 2x NVIDIA B300 or 4x NVIDIA H200 |
| MXFP8 | At least 800 GB | 4x NVIDIA B300 or 8x NVIDIA H200 |
These are data-centre graphics cards, not gaming cards. Thinking Machines recommends the NVFP4 version for all use cases, because the model was trained in that format from the start. Supported software includes SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face's own tools.
Thinking Machines does not list a download size on its own pages. Some reports have given figures in the hundreds of gigabytes; plan for a very large download either way.
How it was trained
The announcement gives unusual detail about training. Thinking Machines says the main pretraining phase used 8.4 trillion tokens and took 68 hours on NVIDIA GB300 NVL72 racks, with the model trained in the NVFP4 format throughout. Further phases stretched the model's context window up to 1 million tokens. Its main reinforcement learning run, the stage where a model is rewarded for good answers and better tool use, ran for nine days on 2,048 GPUs.
The practical takeaway: building a model like this costs far more than any single African startup could spend. Open weights let others reuse that investment rather than repeat it.
Benchmarks, as reported by the company
Thinking Machines published a long list of test scores. These are the company's own figures, and independent testing is still catching up. A few highlights from its announcement:
| Test | What it measures | Score (vendor-reported) |
|---|---|---|
| Artificial Analysis Intelligence Index | Broad mix of reasoning and knowledge tasks | 31 (30.6 in the table) |
| SWEBench Pro v2 (Hard) | Fixing real software bugs | 86.3% |
| ARC-AGI-2 | Abstract puzzle reasoning | 81.7% |
| MMMU Pro | Understanding images and diagrams | 81.6% |
| HLE-Diamond | Very hard expert questions | 28.7% |
ARC-AGI is often discussed in debates about general intelligence; if that term is unfamiliar, our explainer What Is AGI? sets out what it means and why benchmarks alone do not prove it.
Safety notes from the company
Thinking Machines held back the weights for extra safety testing, and it is open about the model's limits. Its announcement says Inkling 2 Small "still trails closed-weights models" on its alignment tests, giving it an internal misbehaviour score of 14.9 against 0.8 for OpenAI's GPT-6.1 Sol. The model card concludes that the model poses no material extra risk beyond what other open-weight models already allow, but that is the developer's own judgement.
The model card also warns of hallucination, weaker performance in long multi-turn conversations, and uneven quality across languages and dialects that were underrepresented in training. It lists English and "general multilingual support" but does not name any African languages.
Why It Matters
Open weights, explained simply
When a company releases "open weights", it lets you download the trained model itself, the huge file of learned numbers, so you can run it on your own computers, change it through further training (fine-tuning) and build it into your own products. That is different from using ChatGPT or Claude, where the model stays on the company's servers and you only send it requests.
Apache 2.0 is one of the most permissive licences available. It lets businesses use, modify and sell products built on the model, with no fee to Thinking Machines, as long as they keep the licence notice. It is not quite the same as fully "open source", because the training data and full training code are not published. Our guide to open-source vs closed AI models explains the trade-offs in more depth.
A new, well-funded player in open models
The open-weight field has been busy this year. We covered Mistral Large 4's open weights and Meta's Muse Spark, and Chinese labs have released many strong open models. Thinking Machines adds a US-based option under a very permissive licence, from a team that includes several well-known former OpenAI researchers. More choice usually means lower prices and fewer single points of failure for developers.
Kept promises build trust
Labs often promise open weights "soon" and then go quiet. Thinking Machines said "next week" and delivered a week later. That matters for anyone deciding whether to build on its models.
What It Means for You
Who can actually run a 594B model?
Realistically, three groups:
- Companies and universities with serious GPU servers. Running the recommended version needs at least four H200-class cards or two B300-class cards working together.
- Anyone renting cloud GPUs. You can rent the same hardware by the hour from cloud providers.
- Everyone else, through hosted providers. Thinking Machines lists API access through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare, plus its own Tinker platform. This is the easiest route for most developers.
A student with a gaming laptop cannot run the full model. Smaller, compressed versions made by the community may appear, but they will be less accurate.
What it could cost in naira
These figures are estimates for illustration only. Prices change often and differ widely between providers and regions.
Renting GPUs. Suppose a cloud provider charges about US$3 per H200 per hour (an illustrative figure, not a quote). Four of them would cost around US$12 an hour, which is about ₦16,800 at roughly ₦1,400 to the dollar. Running them non-stop for a 30-day month would be about US$8,640, or roughly ₦12 million. That is before storage, data transfer and engineering time.
Using a hosted API. AlphaSignal and other reports say Tinker lists Inkling 2 Small at US$0.30 per million input tokens and US$1.00 per million output tokens; Thinking Machines' announcement page does not state a price, so check before you budget. At those reported rates, a small app that processes 20 million input tokens and 5 million output tokens in a month would pay about US$11, roughly ₦15,400.
For most Nigerian startups, the hosted route is far cheaper until usage is very large. If you have never called a model through an API, our guide to building your first app with an AI API walks you through it.
How to try it without spending much
If you want a feel for the model before committing, start small:
- Read the model card first. It sets out the intended uses, the limits and the hardware needed, in fairly plain language.
- Use a playground. Thinking Machines offers a Tinker Playground, and hosted providers often have a web console where you can type prompts without writing code.
- Test with your own material. Use ten or twenty real examples from your work, such as customer emails, reports or code, rather than generic test questions.
- Compare like with like. Run the same examples through the model you use today and note quality, speed and cost.
- Only then think about fine-tuning or self-hosting. Both take time and money, and most teams will not need them at first.
A worked example
Tunde and Halima (an invented example) run a small Abuja startup that summarises tender documents for construction firms. Today they pay a closed-model API. With Inkling 2 Small, they have three options:
- Keep using an API, but switch to a hosted Inkling 2 Small provider. Lowest effort; test whether quality on their documents matches what they have now.
- Fine-tune it on their own past summaries through a provider or Tinker. Because the licence allows commercial use, they can sell the resulting product.
- Self-host later. Only worth it once their monthly API bill is close to the cost of renting GPUs full time, or if a client requires that data never leaves a specific server.
Their sensible first step is a two-week trial on a hosted provider, comparing results on 50 real documents.
Data sovereignty
This is where open weights matter most for Africa. With a closed model, your data travels to the provider's servers, usually abroad. With open weights, a bank, hospital or government agency can, in principle, run the model on servers it controls, including inside Nigeria, so sensitive data never leaves the country.
The catch is hardware. Few data centres on the continent currently offer this class of GPU at scale. That is exactly the gap behind Nigeria's proposal for pooled computing power, which we covered in Nigeria's shared AI compute plan for Africa. A strong, permissively licensed model makes that kind of shared infrastructure more useful.
Language caution
The model card does not mention Yoruba, Igbo, Hausa, Swahili or other African languages, and it warns of uneven quality in underrepresented languages. If your product serves local-language users, test carefully before launch, and consider fine-tuning on your own data.
What to Watch Next
- Independent benchmarks. Expect outside testers to check the company's scores over the coming weeks.
- Smaller community versions. Compressed versions that need less memory are likely; watch their accuracy.
- More platforms. Thinking Machines said the model will reach additional platforms in the coming weeks.
- The funding round. Whether the reported US$5 billion raise is confirmed, and what it funds next.
- Rival releases. The pace is fast; Google's new model is covered in our Gemini 4 Hydrogen report.
Final Thoughts
Inkling 2 Small's open weights are a real addition to the open-model world: a large, capable model under one of the most permissive licences, released when the company said it would be. For most African developers, the practical route is a hosted API, not a home server. But the option to run it on local infrastructure, fine-tune it freely and sell what you build is valuable, especially for data sovereignty. Test it on your own tasks before believing any benchmark, including the company's.
Frequently Asked Questions
What is Thinking Machines Lab?
Thinking Machines Lab is an AI company led by Mira Murati, formerly chief technology officer at OpenAI. It runs Tinker, a platform for fine-tuning models, and builds its own Inkling family of models.
When were Inkling 2 Small's weights released?
Thinking Machines launched the model on 5 November 2026 and updated its announcement on 12 November 2026 to say the weights were downloadable on Hugging Face.
Is Inkling 2 Small free to use commercially?
The weights are released under Apache 2.0, which allows commercial use and modification without a licence fee. You still pay for the computers or hosted service you use to run it.
Can I run Inkling 2 Small on my laptop?
No. The model card says the recommended version needs at least 500 GB of combined GPU memory, such as four NVIDIA H200 cards. Most people will use it through a hosted provider.
How big is Inkling 2 Small?
It has 594 billion parameters in total, with 17 billion active for each token, and handles up to 1 million tokens of context.
Does Inkling 2 Small support African languages?
The model card lists English and general multilingual support but does not name any African languages, so quality in Yoruba, Hausa, Igbo or Swahili is unconfirmed.
Comments (0)
Sign in to post a comment.
- Be the first to comment.


