Every lab releasing an open-weight model right now is chasing the same leaderboard. Thinking Machines Lab, on July 15, 2026, shipped a model that openly admits it is not the strongest one available - and that is the most interesting thing about it.
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling on July 15. Unlike the flagship models from OpenAI, Anthropic, or Google, it is open-weight - outside developers and companies can download and modify it directly.
The company's bet is that enterprises are less interested in the most capable general-purpose AI on a leaderboard than in a model they can genuinely make their own.
That is a specific, falsifiable claim. Let's see if the architecture backs it up.
What Inkling actually is
Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio, and video.
The active-parameter figure is the one that matters operationally. Inkling has 975 billion total parameters but only draws on 41 billion active parameters during any given request, which keeps compute costs manageable without sacrificing the model's breadth. That ratio - roughly 4% active - is consistent with what teams actually need to self-host a model of this quality without dedicated cluster hardware.
Inkling is released under Apache 2.0 , which matters more than most coverage notes. Apache 2.0 imposes no usage restrictions and no attribution requirements in outputs. You can build a commercial product on it, modify it, redistribute it, and never mention Thinking Machines Lab. Mistral's Large 3 is also Apache 2.0; most other frontier-adjacent models are not.
Alongside Inkling, Thinking Machines Lab is sharing a preview of Inkling-Small, a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency.
Where it sits on benchmarks - and where it doesn't
On SWE-bench Verified, Inkling scores 77.6% - ahead of Nemotron 3 Ultra (70.7%), roughly level with Kimi K2.6 (80.2%) and GLM 5.2 (80.0%), and behind closed models like Claude Fable 5 (95.0%). On reasoning, GPQA Diamond comes in at 87.2% and AIME 2026 at 97.1%.
On Terminal-Bench 2.1, Inkling scores 63.8% - mid-pack among open weights; it trails GLM 5.2 (82.7%) there.
The non-obvious thing here: Thinking Machines is not hiding these gaps. The release materials name them. That is a design choice. A model positioned as a fine-tuning base does not need to win every benchmark out of the box - it needs to be a solid, steerable starting point that improves predictably when you train on your own data. General-purpose SOTA and domain-adapted SOTA are different targets, and conflating them is how teams end up paying for capability they never use.
The fine-tuning-first design choice
Thinking Machines trained Inkling for calibration, instruction following, and resistance to censorship - what the lab calls the model's epistemics. Getting the facts right requires more than memorizing a large corpus of knowledge. A useful model must be well-calibrated, expressing the right amount of confidence in its answers, including on questions that are not yet settled.
That calibration framing is unusual. Most labs optimize for benchmark peaks. Calibration means the model's confidence tracks its actual accuracy - it says "I don't know" when it doesn't know, rather than confabulating with false certainty. For a base model that teams plan to fine-tune on proprietary data, this matters: bad calibration is hard to fix in post-training and poisons downstream trust.
On FORTRESS adversarial safety, Inkling scores 78.0% - the highest built-in safeguards of any open-weight model the lab compared, without over-refusing benign look-alike queries. Over-refusal is a real cost for enterprise deployments; a model that hedges on every borderline prompt creates friction that teams route around.
Thinking Machines is not monetizing Inkling through metered API access. The weights are free. The monetization model is presumably Tinker - the fine-tuning platform - and future products. That aligns the lab's incentives with making the base model genuinely good to customize, not with locking you into inference fees.
Beagle in action#ai-ops, 11:03amWhat this means compared to the current open-weight wave
Two other major open-weight releases landed in roughly the same window. On July 16, Moonshot shipped Kimi K3, a 2.8-trillion-parameter open-weight mixture-of-experts model with native vision and a 1-million-token context window. It is the largest open-weight model released to date, and early independent reviews place it at frontier level for agentic coding.
Benchmark tests show the open models are nearly as good as top American frontier models, but at just a fraction of the cost.
Inkling is deliberately not competing on that axis. Here is the comparison that matters for a team making an actual build decision:
| Inkling | Kimi K3 | GLM-5.2 | |
|---|---|---|---|
| Release date | July 15 | July 16 | June 13 |
| Total params | 975B | 2.8T | 744B |
| Active params | 41B | ~16/896 experts | 40B |
| License | Apache 2.0 | Modified MIT | MIT |
| SWE-bench Verified | 77.6% | frontier-level† | 80.0%‡ |
| Native multimodal | yes | yes | no |
| Fine-tuning tooling | Tinker (first-party) | none announced | none announced |
| US-built | yes | no | no |
†Kimi K3 full open weights scheduled July 27; ‡GLM-5.2 SWE-bench score is on SWE-bench Pro - not directly comparable.
Inkling is a notable US-built entry in a field where many open-weight options come from China. That matters for teams in regulated industries or with data-residency requirements who have been watching the Kimi/GLM wave with interest but cannot route data through Chinese infrastructure.
The real question: do you have data worth fine-tuning on?
Inkling's design makes one thing obvious - the value of an open-weight model built for customization depends entirely on whether you have domain data worth using. The model gives you a high-quality, well-calibrated base. The fine-tuning toolchain (Tinker) lowers the operational bar. But if your team's actual domain knowledge lives scattered across Notion pages, Slack threads, and undocumented tribal knowledge, fine-tuning won't help much. That is the real prerequisite, and it is worth auditing before you invest in the infrastructure.
A teammate like Beagle, living in Slack, can surface and consolidate the in-thread knowledge that rarely makes it into training data - which makes it useful upstream of any fine-tuning decision, not just downstream.
Open-weight model fine-tuning: common questions
What is Inkling and who made it?
Inkling is a 975B-parameter open-weight multimodal model released July 15, 2026, by Thinking Machines Lab - the startup founded by former OpenAI CTO Mira Murati. It uses a Mixture-of-Experts architecture with 41B active parameters, a 1M-token context window, and an Apache 2.0 license. It is designed as a fine-tuning base, not a leaderboard topper.
How does Inkling compare to Kimi K3 and GLM-5.2?
Kimi K3 is larger (2.8T parameters) and targets outright agentic coding performance. GLM-5.2 leads on long-horizon coding benchmarks. Inkling scores 77.6% on SWE-bench Verified - mid-pack among current open weights - but is the only one in this tier with a first-party fine-tuning platform (Tinker) and Apache 2.0 licensing that imposes no commercial restrictions.
Can I fine-tune an open-weight model on my company's data?
Yes, and Inkling is designed for exactly this. The Apache 2.0 license permits modification and commercial redistribution. Thinking Machines Lab ships Tinker alongside Inkling specifically for fine-tuning. The practical prerequisite is clean, domain-specific data - most teams underestimate how long data preparation takes relative to the actual training run.
Is Inkling safe to use in production?
Thinking Machines commissioned external safety testers before release. Inkling scored 78.0% on the FORTRESS adversarial safety benchmark - the highest among open-weight models in the lab's own comparison set - with low over-refusal on benign queries. As with any open-weight model, you are responsible for your deployment's safety guardrails; the base weights are a starting point, not a compliance certificate.
What does Apache 2.0 mean for an open-weight model?
Apache 2.0 lets you use, modify, distribute, and commercialize the model without royalties or output attribution requirements. It is one of the most permissive licenses in the open-weight space. By contrast, some open-weight models carry modified licenses that restrict commercial use, prohibit certain deployments, or require attribution. Check the fine print - the license difference between "open weight" models can be larger than the benchmark difference.