Open-weight models grew commercial strings this summer

The summer's open-weight wave arrived with fine print. Between July 16 and August 13, Moonshot AI shipped Kimi K3, Alibaba shipped Qwen3.8-Max, DeepSeek released V4-Pro, and Meta returned to open weights with Muse Glimmer. All four publish downloadable weights. But the terms attached to those weights now run from plain Apache 2.0 to custom licenses with revenue-share clauses — which means license diligence has quietly become part of model selection.

Four releases, four sets of terms

Moonshot AI announced Kimi K3 on July 16: a mixture-of-experts model with 2.8 trillion total parameters (104 billion active per token), multimodal input, and a 1M-token context window, marketed as the first open 3T-class model. Weights followed July 27 under a modified MIT license. Artificial Analysis measured an overall Elo of 1547 on its knowledge-work evaluation, up 732 points from Kimi K2.6. API pricing is $3 per million input tokens and $15 per million output — the most expensive release yet from a Chinese lab. And per AI News reporting, Moonshot's commercial framework can require partners to share up to 30% of revenue once combined revenue from the model exceeds $20 million over 12 consecutive months.

Alibaba's Qwen3.8-Max launched August 3: a 2.4-trillion-parameter MoE (95 billion active) with text, image, and video input, a 1M-token context window, and API pricing of $2 per million input and $6 per million output. Open weights went live August 12 under a custom "qwen3.8-max" license — the first Max-class Qwen flagship to be open-weighted — with reporting indicating a revenue-share clause for large commercial users. Two details matter for buyers: the BF16 weights are roughly 4.89 TB, and the open variant is not the hosted model. It is text-only, lacks native 1M context, and requires thinking mode.

DeepSeek posted V4-Pro weights to Hugging Face on August 12 under plain MIT — roughly 1.6 trillion total parameters, 49 billion active, an 893 GB download — and made the model generally available on August 13 with a 1M-token context and outputs up to 384,000 tokens. Then it repriced the API. Effective August 16, peak-hour output rises to $3.96 per million tokens from a $0.87 flat rate, with off-peak at half the peak price; coverage citing the change log noted increases of up to ~14× on some tiers.

Meta went the other way. Muse Glimmer shipped August 10 under Apache 2.0 with ungated weights on Hugging Face and no Llama-style user cap, reversing the closed strategy Meta set when Muse Spark launched proprietary-only in April. It is a 30B dense multimodal model distilled from Muse Spark and tuned for multi-step agent workflows, with a 131,072-token context. Meta reports MCP Atlas 75.5 and SWE-Bench Pro 51.2, and says a quantized build runs in 24GB of VRAM on an RTX 3090 or 4090.

What an open-weight license actually allows

The word "open" is doing less work than it used to. Apache 2.0 and MIT releases carry no commercial strings: you can self-host, fine-tune, embed, and resell. Modified licenses are different instruments. A revenue-share trigger like Moonshot's turns a model download into a commercial contract with a $20 million tripwire. A custom license like Alibaba's requires contract review before you build a product or platform on the model — and the reported terms target exactly the companies most likely to standardize on it: large commercial users offering the model as a service.

Open weights tell you what you can download. The license tells you what you can build.

None of this makes the restricted releases bad options. It makes them negotiated options. The practical change is procedural: license review moves from a formality to a gating step in model selection, alongside benchmarks and unit economics.

Self-hosting vs API pricing: run the numbers

DeepSeek's repricing is the clearest signal in the window. API pricing is a dial the vendor can turn — up as well as down — and a cost model built on $0.87 output is about to become a cost model built on $3.96 at peak. The MIT-licensed weights are a genuine hedge, but an 893 GB, 1.6T-parameter model is a serious infrastructure commitment: real for large IT shops, theoretical for most mid-market teams. Kimi K3, at 2.8 trillion parameters, is effectively hosted-access-only for most enterprises.

Muse Glimmer changes the small end of that comparison. A 30B Apache 2.0 agent model on a single 24–32GB GPU fits existing workstation and edge budgets, which makes self-hosted pilots cheap to run — particularly for offline or data-sensitive agent workloads where per-token APIs were never a comfortable fit. Meta measures its speculative-decoding drafter at up to 3.1× faster generation on an RTX 5090, with 0.2%–1% accuracy degradation from quantization; validate both claims on your own tasks.

What operators should do now

• Read the license before the benchmark table. Classify every candidate as permissive (Apache 2.0, MIT) or negotiated (modified, custom) before evaluation effort is spent.
• Route revenue-share terms to legal. Thresholds like $20 million over 12 months sound distant until a product ships. Get the trigger conditions in writing.
• Confirm the open variant matches the hosted one. Qwen3.8-Max's downloadable weights lack the hosted API's vision input and 1M context. Test the artifact you would actually deploy.
• Price API volatility into steady workloads. DeepSeek's roughly 4.5× peak increase is the case study; off-peak scheduling for batch jobs is a real lever.
• Pilot a small permissive model. A 30B Apache 2.0 model on one GPU is a low-cost way to learn what self-hosting demands of your team.

The license spectrum sharpens a question we work through with clients constantly: when to buy hosted capability, and when to build on weights you control. Our framework in build vs buy for AI agents applies directly, and this month's hosted inference price drops are the other half of the comparison. If you are selecting a model for a steady production workload, book a consult.