Thinking Machines Inkling: open-weight model for SMBs
Mira Murati's company Thinking Machines Lab released on 15 July 2026 its first open model, called Inkling. 975B parameters total, 41B active per token, up to 1M tokens context window, and fully open under Apache 2.0 (Thinking Machines). For a Norwegian SMB leader evaluating AI vendors this means three concrete things: you can download the model weights, you can run them in the EU without sending data to the US, and you avoid the expensive dependency on a single API provider that can change its price tomorrow.
At the same time we have to be honest. Inkling is not the strongest model on the market. Thinking Machines itself says: "It is not the most performant model available today, closed or open." On HLE text only Inkling scores 30 percent, while Claude Fable 5 scores 53.3 percent (Thinking Machines benchmarks). It is built for solidity and openness, not for beating frontier models on every test. That is a deliberate choice, and for many Norwegian SMBs it is exactly what they need.
What is actually new about Inkling?
Inkling is a MoE transformer (mixture-of-experts) with 66 layers, where each token activates 6 of 256 experts, plus 2 shared experts. That means you get the capacity of a 975B model, but the compute cost of a 41B model per query (Thinking Machines model card).
Three things set it apart from most other open models in 2026:
- Multimodal from the ground up. Inkling is trained on 45 trillion tokens of text, image, audio and video in the same pre-training run. It uses an encoder-free architecture for audio (discrete dMel spectrograms) and 40x40 image patches via a four-layer hMLP. For Norwegian SMBs that want to process customer call recordings, product images and documents in the same workflow, this is relevant.
- Variable thinking depth. You can adjust how much the model "thinks" per task, from 0.2 to 0.99. On Terminal Bench 2.1 Inkling uses one third as many tokens as Nemotron 3 Ultra for the same result (Thinking Machines). That directly affects price and response time.
- Stronger epistemics. Inkling is trained to be calibrated on forecasts, not overconfident in itself. On ForecastBench without search it scores 61.1 percent, better than GPT-5.5 at 59.1 and Claude Opus 4.8 at 54.6. For case handling and decision support this matters more than raw reasoning score.
Why is this interesting for Norwegian SMBs?
The most important reason is control. With a closed model via API you are always a guest in someone else's house. They can change the price, they can change the terms, they can change which model actually answers behind the scenes. We have seen this several times in recent months, including in the ChatGPT Work data deletion case and in Anthropic's agent SDK pause in June.
Apache 2.0 plus open weights flip this around. You own the deployment. You can run it with an EU host, you can finetune it on your own data, and you avoid sending business-critical information through an API where you do not know how the data is used.
For Norwegian SMBs it is especially relevant to see this as a Western open-weight alternative to Chinese models. DeepSeek V4 Pro and Qwen are technically strong and popular in the open-weight community, but for a Norwegian business with customer data, public procurement contracts or defence-related customers the country of origin of the weights is a central part of the assessment. Inkling is trained in San Francisco, on NVIDIA GB300 NVL72, and the weights live on Hugging Face.
What does it take to run Inkling?
Here comes the less cheerful news. Full precision (BF16) requires around 2 TB of aggregated VRAM. With NVFP4 quantisation the requirement is around 600 GB. That is hardware that costs millions of NOK to acquire and operate, and it is out of reach for the vast majority of Norwegian SMBs.
Warning light for SMBs: Full self-host of Inkling in 2026 is a project for a large technology department or a specialised hosting provider, not for an accounting firm with eight staff.
The realistic path in for an SMB is one of three:
- Tinker API for finetuning. Tinker is Thinking Machines' training platform, released in October 2025. You send data, they train a customised version of Inkling for you, and you get a model you can use. Pricing has not been publicly announced yet, but it is separate from the inference pricing (Thinking Machines).
- Hosted inference. Several providers will offer Inkling as an API service in the coming weeks, including through SGLang, vLLM, TokenSpeed and Unsloth. For an SMB that wants to use the model without running it themselves, this is the entry point.
- EU-hosted on-prem via partner. For SMBs in regulated industries (healthcare, finance, public sector) a Norwegian or Nordic GPU hosting partner may be the right path. Then you get control over the data without buying the GPU cluster yourself.
What can Inkling be used for in practice?
We see four concrete use cases where Inkling fits well for Norwegian SMBs, and where the open licence actually makes sense:
- Finetuning on Norwegian domain data. An accounting firm can train an Inkling variant on Norwegian accounting standard (NGAAP), a law firm on Norwegian legal practice, a healthcare provider on Norwegian patient record language. This is what we at AIKI help customers do through AI Kickstart, where we build a production-ready workflow in 1-4 weeks at a fixed price.
- Agent systems with controlled data access. Inkling is trained to be good at agent work, and trained with varied tools that work in practice. For SMBs building autonomous agents Inkling is an alternative to the expensive closed models.
- Multimodal document flow. Inkling handles text, image, audio and video in the same model. For SMBs that want to build a customer service that reads PDF, listens to recordings and answers in chat, this is a simplification compared to running four specialised models.
- Forecasts and decision support. Inkling scores high on calibrated forecasts, better than GPT-5.5 and Claude Opus 4.8. For SMBs working on demand forecasting, price analysis or risk evaluation this is a concrete use case.
Three moves AIKI recommends this week
1. Map one workflow where you send sensitive data to an API today. Do not switch everything at once. Pick one concrete case, for example contract analysis or customer log analysis, and assess whether Inkling via an EU host can replace the API call. We at AIKI do this in 1-2 weeks through an AI Review, where we also look at total cost per resolved task.
2. Test Inkling against the model you use today on your own data. Once hosted inference is available in Norway (in the coming weeks), run an A/B test. Measure not just quality, but also cost per task. A model that scores 10 percent lower, but costs half, can be a better choice for high-volume jobs. We have previously written about why the AI bill is exploding for many SMBs, and this is a concrete antidote.
3. Set a policy for when open-weight is the right choice. Do not use Inkling for everything, and do not use it for nothing. Make an internal rule: tasks with sensitive data, Norwegian domain language or customisation needs go through an open model. Standard tasks that do not require special customisation can still use the best closed models. This is the same logic we recommend for AI pricing and model choice for Norwegian SMBs.
What Inkling is not
To be clear, as Thinking Machines is itself:
- Not SOTA on reasoning. 30 percent on HLE text only versus Fable 5 at 53.3 is a large gap. For demanding codebase migration or long, complex document analysis, closed frontier models are still a better choice.
- Not free to run yourself. The infrastructure cost is high. What is free are the weights and the licence, not the operations.
- Not a substitute for compliance work. An open model does not solve GDPR, privacy or data processing agreements. You still need to control the data flow.
Summary
Inkling is not the model that wins the benchmark table. It is the model that gives Norwegian SMBs a real alternative to building the entire AI strategy on one American provider's API. For an SMB leader who spent 2025 wondering whether AI is worth the money, and who now wants more control over both cost and data sovereignty, this is a mature tool to start testing.
The most important change Inkling brings is not technical. It is that open models in 2026 are good enough for production in many SMB scenarios, and that Apache 2.0 plus available weights gives you a negotiating position you do not have when you are one hundred percent locked to one API.
We are tracking when hosted inference in the EU region becomes available for Norwegian customers, and we will update the article as soon as we have concrete numbers from a Norwegian production environment.
Want to know how a specific workflow in your business can be moved to an open model without losing quality? Book a non-binding conversation with us, or start with an AI Review to map out where open models actually pay off in your operations.
Frequently asked questions about Thinking Machines Inkling
What is Thinking Machines Inkling?
Inkling is an open-weight multimodal LLM from Thinking Machines Lab, founded by Mira Murati. The model was released on 15 July 2026 with 975B total parameters and 41B active per token, a 1M token context window and an Apache 2.0 licence. The weights are available on Hugging Face.
Is Inkling as good as GPT-5.6 or Claude Fable 5?
No. Thinking Machines is itself clear that Inkling is not the strongest model on the market. On most benchmarks it scores below Claude Fable 5 and GPT-5.6 Sol. The point of Inkling is that you get a solid open model you can own and customise, not that it should beat the best closed models.
Can a Norwegian SMB run Inkling on its own premises?
Not at full scale. Full precision requires around 2 TB of aggregated VRAM. NVFP4 quantised, the requirement is around 600 GB. That is infrastructure that costs millions. The realistic path for an SMB is the Tinker API for finetuning, or hosted inference with an EU provider.
What does Apache 2.0 mean in practice?
Apache 2.0 gives you a commercial usage right, the right to modify, the right to redistribute, and a patent clause that protects you against patent claims from contributors. For a Norwegian SMB building a product with AI at the core it means you can use Inkling as a foundation without paying licence fees or sharing revenue.
Why is Inkling interesting for Norwegian SMBs if it is not the best?
Three reasons. You avoid being locked to a single API provider that can change price or terms. You can finetune on Norwegian domain data without sending it to an American cloud. And you get a Western open-weight alternative to Chinese models like DeepSeek V4 Pro and Qwen, which many Norwegian organisations are sceptical of for data sovereignty reasons.
What is Tinker and why is it relevant?
Tinker is Thinking Machines' finetuning platform, released in October 2025. It lets you train customised versions of Inkling on your own data without running a GPU cluster yourself. For SMBs that want to build a Norwegian specialist model for contract analysis, case handling or customer dialogue, this is the most concrete entry point. /usr/bin/bash: line 5: /tmp/hermes-snap-492b4dcf67d7.sh: No such file or directory /usr/bin/bash: line 6: /tmp/hermes-cwd-492b4dcf67d7.txt: No such file or directory /usr/bin/bash: line 5: /tmp/hermes-snap-492b4dcf67d7.sh: No such file or directory /usr/bin/bash: line 6: /tmp/hermes-cwd-492b4dcf67d7.txt: No such file or directory /usr/bin/bash: line 5: /tmp/hermes-snap-492b4dcf67d7.sh: No such file or directory /usr/bin/bash: line 6: /tmp/hermes-cwd-492b4dcf67d7.txt: No such file or directory /usr/bin/bash: line 5: /tmp/hermes-snap-1e4063f7f7cb.sh: No such file or directory /usr/bin/bash: line 6: /tmp/hermes-cwd-1e4063f7f7cb.txt: No such file or directory /usr/bin/bash: line 5: /tmp/hermes-snap-25f3ad572298.sh: No such file or directory /usr/bin/bash: line 6: /tmp/hermes-cwd-25f3ad572298.txt: No such file or directory /usr/bin/bash: line 5: /tmp/hermes-snap-25f3ad572298.sh: No such file or directory /usr/bin/bash: line 6: /tmp/hermes-cwd-25f3ad572298.txt: No such file or directory
