Thinking Machines Lab shipped Inkling on July 15, 2026. It runs 975 billion parameters. It ships under Apache 2.0. And it is the best open weights any US lab has put out.
That last sentence about open weights is doing a lot of quiet work. Inkling debuted at 41 on the Artificial Analysis Intelligence Index. Kimi K3 sits at roughly 57. GLM-5.2 sits at 51.
So the leading American open weights release lands third or fourth in its own category. The interesting question is not whether Inkling wins. It is what a 975B model with a permissive license actually changes for anyone downstream.
The honest answer: less than the launch suggests, and not what most coverage claims.
Key Takeaways on Frontier-Scale Open Weights
- Inkling is a 975B-parameter MoE with 41B active per token, a 1M-token context window, and Apache 2.0 terms.
- Running it at BF16 needs roughly 2TB of aggregated VRAM. Downloadable does not mean runnable.
- Six labs shipped open weights above 100B in 2026. Five of them are Chinese.
- licenses diverged sharply this year. Kimi K3 dropped Modified MIT for a bespoke document with a $20M revenue gate.
- The real shift from open weights is control over deployment, not access to capability.
Quick Navigation
- What Inkling 975B Actually Is
- Every Open-Weight Release Above 100B in 2026
- What Open Weights Actually Change
- What Open Weights Do Not Change
- What Inkling-Small Changes
- The Fine-Tuning Economics Behind Open Weights
- Why Five of Six Frontier Open Weights Are Chinese
- The Safety Argument Around Open Weights
- How to Choose Among 2026 Open Weights
- Conclusion: Open Weights Changed the Contract, Not the Compute
- What This Table Will Look Like in Six Months
- FAQ About Inkling and 2026 Open Weights
What Inkling 975B Actually Is
Start with the specification, because the shape explains why these open weights exist.
Inkling is a Mixture-of-Experts transformer: 975B total parameters, 41B active per token. It runs 66 decoder layers with 256 routed experts plus 2 shared experts, and routes each token to 6 of the routed set. If those terms are unfamiliar, our glossary covers MoE and context windows.
Pretraining used 45 trillion tokens of text, images, audio, and video. Training ran on NVIDIA GB300 NVL72 systems, using Muon for large matrix parameters and Adam for the rest.
Two design choices stand out in the architecture notes. Short convolutions inside the attention block give an explicit path for mixing nearby tokens. A separate RMSNorm sits directly after the embedding lookup.
Neither ships with an ablation, so their contribution is unmeasured.
The third choice is the practical one. Most post-training compute went to asynchronous reinforcement learning past 30 million rollouts, and that run produced a controllable effort dial. You set reasoning_effort and the model spends a different token budget.
The launch post is unusually candid. The company states plainly that Inkling is not the strongest model available today, open or closed.
That framing matters. The pitch is breadth and fine-tunability, not a leaderboard position. Open weights here are a distribution strategy, not a capability claim.
Every Open-Weight Release Above 100B in 2026
Here is the full open weights field, ordered by ship date. Parameter shape and license are the two facts that do not go stale in a week.
| Model | Lab | Released | Total / Active | Context | Licence |
|---|---|---|---|---|---|
| Kimi K2.6 | Moonshot AI (CN) | 20 Apr 2026 | 1T / 32B | 256K | Modified MIT |
| DeepSeek V4 | DeepSeek (CN) | 24 Apr 2026 (preview) | 1.6T and 284B variants | 1M | MIT |
| Mistral Medium 3.5 | Mistral (FR) | 29 Apr 2026 | 128B dense | — | Modified MIT |
| MiniMax M3 | MiniMax (CN) | 1 Jun 2026 | Not disclosed | 1M | Not verified |
| Kimi K2.7 Code | Moonshot AI (CN) | 13 Jun 2026 | ~1T / 32B | — | Disputed (see below) |
| GLM-5.2 | Z.ai / Zhipu (CN) | mid-Jun 2026 | 744B / ~40B | 1M | MIT |
| Inkling | Thinking Machines (US) | 15 Jul 2026 | 975B / 41B | 1M | Apache 2.0 |
| Kimi K3 | Moonshot AI (CN) | weights 26–27 Jul 2026 | 2.8T / 104B | 1M | Custom “Kimi K3 License” |
Caveats on the table
Three entries need flags, and no other roundup I found carries them.
MiniMax M3. Parameter count is not publicly specified in the sources I could verify. It is included on the strength of frontier positioning, not a confirmed figure.
Kimi K2.7 Code. Sources disagree on license. Some list Apache 2.0, others list the Modified MIT that governs the K2 line. Check the model card before building on it.
Kimi K3 dates. Hosted launch was 16 July. Weights landed 26 July, one day ahead of the stated 27 July target. Both dates appear in coverage.
Two exclusions from the open weights table
Gemma 4 is excluded from the open weights table because its largest variant is 31B dense. Llama 4 Scout is excluded as a 2025 release.
What the Licence Column Tells You About Open Weights
Read that table by license rather than parameter count and a different picture appears.
Apache 2.0 and MIT are permissive licences, not unconditional ones. They permit commercial use without a revenue or user-count gate, but Apache 2.0 and MIT have notice and redistribution conditions. Check the actual licence shipped with each model rather than treating “open weights” as a single legal category.
That is the strongest thing about the Inkling release. Among frontier-scale open weights, Apache 2.0 with no attached usage policy is the cleanest set of terms on offer.
Modified MIT sounds permissive, and mostly is. Still, open weights with a clause are not weights without one. The K2-family clause requires prominent “Kimi K2” attribution once a product passes 100 million monthly active users or $20 million in monthly revenue.
Most teams will never hit that. But it is a term, and terms compound across a stack.
Kimi K3 broke the open weights pattern. Moonshot replaced Modified MIT with a bespoke license, and the change went largely unexamined in launch coverage.
Simon Willison flagged the K2 lineage, and the K3 document goes further. Reports say firms above $20 million in annual revenue must sign a contract with Moonshot before offering K3 to outside customers as a service. Attribution rules apply on top.
So the largest open weights model in the world is not open by the Open Source Initiative definition. Neither is it uniquely restrictive. It is a commercial license wearing an open label. So read it with a lawyer, not a skim.
What Open Weights Actually Change
Three things genuinely shift when open weights ship publicly. None of them is “everyone can now run frontier AI.”
Open weights give you deployment control
Open weights let you choose where the model runs. That is the whole thing, and for regulated buyers it is enormous.
A hospital, a bank, or a defense firm can keep the model inside its own walls. No data leaves. The vendor never sees a prompt, and API terms cannot change under you mid-contract.
Price discipline
So open weights cap what closed vendors can charge for the same capability. When GLM-5.2 delivers similar coding performance at a fraction of frontier pricing, that becomes the reference point in every procurement conversation.
Still, the effect reaches teams who never self-host. They simply negotiate better, because open weights set the floor.
Modification rights
Fine-tuning on your own data is the pitch behind Inkling. Tinker exists to make that path short, and a broad base model adapts to more workflows than a narrow one.
Distillation matters here too. Permissive weights can support developing a smaller student model, subject to the applicable licence conditions and any other rights involved. Apache 2.0 should not be treated as a blanket promise that every resulting asset is owned free of third-party obligations.
What Open Weights Do Not Change
Now the correction, because open weights coverage overclaims here.
Open weights access is still gated by hardware
Inkling open weights need roughly 2TB of aggregated VRAM at BF16. NVFP4 quantization cuts that substantially but requires SM100-class hardware.
Kimi K3’s checkpoint runs 1.56TB across 96 shards. Self-hosting it realistically means eight to sixteen nodes of eight H100 or B200 accelerators.
So “open” in open weights describes the license, not the barrier. Memory and interconnect economics still set the ceiling, and those have not moved because a download link appeared.
Community quantization helps at the edges. One 1-bit GGUF cut Kimi K3 from 1.56TB to 594GB, keeping about 79% accuracy. Still a serious machine, though.
Reproducibility is not included
Thinking Machines says open weights rather than open source, and the distinction is precise. Training data and the training pipeline stay private.
Every release in that table does the same. So you get the artifact, never the recipe. That is the hard limit on what open weights can prove. So you cannot audit what went in, verify contamination claims, or rebuild the model from scratch.
Benchmarks still need care
Inkling’s Terminal Bench 2.1 figures come from an internal harness, while competitor scores are self-reported. Those are not directly comparable.
It also trails GLM-5.2 and Kimi K2.6 on HLE, Terminal Bench, and SWE-bench Verified, and posts 43.9% on SimpleQA Verified against DeepSeek V4 Pro’s 57.0%. We covered why leaderboard gaps at this level are hard to interpret.
One number does stand out. Inkling posts the highest FORTRESS adversarial score among compared open-weights models at 78.0%, which matters more for regulated deployment than another point of coding accuracy.
What Inkling-Small Changes
The open weights nobody can download yet may matter more than the ones that shipped.
Thinking Machines Lab released Inkling-Small and its full weights on 30 July 2026. It has 276B total parameters and 12B active. The released model’s model card should be used for specifications and evaluation settings, rather than carrying forward preview benchmark numbers.
Read that twice. So the small model wins on some benchmarks, while the big one carries the headline.
Why size beats score here
Active parameters describe the computation used per token, not the complete weight-memory requirement. Inkling-Small still has 276B total parameters. At BF16, weights alone imply roughly 552 GB before runtime overhead; quantization and offloading change that estimate. Hardware suitability therefore depends on precision, memory, context and serving configuration, not the 12B-active label alone.
The full Inkling-Small weights are now published. The practical access question has moved from release timing to the memory, throughput and operating-cost tradeoffs of a real deployment.
The Fine-Tuning Economics Behind Open Weights
Thinking Machines is not really selling open weights. It is selling a customization pipeline.
Tinker exists to make fine-tuning short. So Inkling was trained broadly rather than narrowly, because open weights only pay off if people adapt them. Breadth adapts to more workflows than a specialist base does.
Fine-tuning a frontier-scale model on your own data is expensive, and the result is yours. Calling a closed API is cheap per token, and the result is rented.
Open weights change which side of that trade is available. But switching costs rise once you adapt a model. So the lock-in moves rather than vanishing.
Why Five of Six Frontier Open Weights Are Chinese
The geography of frontier open weights is the most under discussed fact in that table.
Moonshot, DeepSeek, Z.ai, and MiniMax all ship at this scale routinely. American labs mostly do not, and Inkling is notable partly because it breaks a pattern.
Open weights are a share-capture move when you are behind on distribution. A downloadable model gets into stacks that would never sign an API contract with a Chinese vendor.
Export controls push the same way. If you cannot match a rival’s compute budget, giving the weights away buys reach instead. Reach compounds differently than revenue does.
But adoption of open weights is not purely technical. Moonshot has faced accusations, including from the White House OSTP director, that K3 was distilled from a competitor’s model. Those claims are unresolved.
Regulated US buyers weigh provenance alongside benchmarks. That is the gap Inkling aims at, even with a lower index score.
The Safety Argument Around Open Weights
Publishing open weights is irreversible, and that drives most of the disagreement.
Once a checkpoint is out and mirrored, no vendor can pull it back or patch it. And anyone with modest compute can fine-tune the safety training away.
Critics argue frontier open weights hand capability to actors who could not build it. That concern centres on cyber and biological uplift, and it does not depend on the license at all.
Supporters point out that inspection requires access. Outside researchers cannot audit a model they can only query through a filtered API.
The MarkTechPost breakdown notes Thinking Machines flags role-play and indirect prompts as residual risks in its own project page. That kind of published limitation is only possible when someone can test for it.
Evidence has not settled either side, so this post will not settle it either. But the FORTRESS score in the Inkling release suggests labs are starting to compete on adversarial robustness, which is a healthier signal than benchmark parity.
How to Choose Among 2026 Open Weights
Skip the leaderboard for a moment. Four questions decide most open weights selections.
What can you actually run? Start with available VRAM, then filter the open weights list. A 744B model you can serve beats a 2.8T model you cannot.
What does the license require at your scale? Check revenue and user thresholds against your projections, not your current numbers. MIT and Apache 2.0 have neither.
Do you need the weights, or just the price? If you will call an API anyway, open weights matter to you only as negotiating leverage.
How much does provenance matter? For some buyers it decides everything, and no benchmark will move them. Open weights from a US lab answer a question a score cannot.
Conclusion: Open Weights Changed the Contract, Not the Compute
Inkling is a good model with excellent terms. Yet it will not top a leaderboard, and its makers said so first. Thinking Machines said so themselves, which is more than most launches manage.
The significance of these open weights sits elsewhere. The 2026 field now offers genuine frontier-scale capability under Apache 2.0 and MIT, which was not true two years ago.
But the constraint moved rather than disappearing. Access to weights is now free. Access to the two terabytes of memory needed to serve them is not, and that gap decides who actually benefits.
Watch deployment results for Inkling-Small now that its weights are available: memory use, useful throughput and cost at the required quality. Separately, compare the actual commercial and redistribution terms of each model licence rather than treating different open-weight releases as legally interchangeable.
What This Table Will Look Like in Six Months
Two forces will reshape it, and they pull opposite ways.
Scale keeps climbing. Kimi K3 crossed the 3-trillion class, so the next tier arrives before year end. Yet each jump narrows the pool of buyers who can serve the result.
Smaller active-parameter counts can reduce compute per token. Adoption still depends on total weight storage, usable throughput and the workload, so both active and total parameters belong in the comparison.
So expect this open weights table to split in two. One row set for labs proving capability, another for models people genuinely run.
FAQ About Inkling and 2026 Open Weights
What licence does Inkling use?
Apache 2.0, with no attached usage policy. That permits commercial deployment, fine-tuning, and redistribution without royalty or revenue thresholds. Thinking Machines describes the release as open-weights rather than open source, because training data and the training pipeline are not published.
What hardware do you need to run Inkling 975B?
Running the open weights takes roughly 2TB of aggregated VRAM at BF16. NVFP4 W4A4 quantization reduces that considerably but requires SM100-class hardware or newer. In practice this means a multi-node GPU cluster, not a workstation.
Is Inkling better than Kimi K3 or GLM-5.2?
Not on aggregate benchmarks, though license terms differ. Inkling debuted at 41 on the Artificial Analysis Intelligence Index against roughly 57 for Kimi K3 and 51 for GLM-5.2. It leads on FORTRESS adversarial robustness at 78.0% and carries the cleanest license of the three.
Which 2026 open weights have the most permissive license?
Inkling uses Apache 2.0. Apache 2.0 and MIT permit commercial use without revenue/user gates, but include notice and redistribution conditions. They should not be described as having no attribution requirements. Verify the exact current licence for each other model before relying on the comparison.
Does open weights mean open source?
No. Open weights means the trained parameters are downloadable. Open source, under the Open Source Initiative definition, additionally requires the training data and code, and a license without discriminatory conditions. Every frontier-scale release in 2026 publishes weights only.
Why do Chinese labs release open weights more often?
Distribution strategy under constraint drives most open weights releases. Open weights get a model into stacks that would not sign a vendor contract, and they convert compute-limited capability into ecosystem position. Export controls make that trade more attractive than competing on API revenue alone.
Keep reading
Here are the latest posts from the blog.
LLM Context Window Comparison 2026: Limits and Costs
LLM Inference Latency by Model and Provider: Real Numbers
GPU Cloud Pricing 2026: What an Hour Actually Costs
