Feeling concerned about new developments of LLMs in terms of access and openness. Anthropic is gating access to Mythos, could this point to a future were frontier models are only for sanctioned partners? Coincidentally, Qwen is releasing their best current model through closed API only (currently ranking 5th in the openhands index). Meanwhile, MiniMax-M2.7 is open but with a non-commercial license, a departure from M2.5 (interestingly, it’s ranking below M2.5 at the openhands index in the overall category). Also wondering what is keeping DeepSeek from making a new release. Their last release is from Dec 2025. Any thoughts?

nod OpenAI released many models only via their apps for weeks+, before being in the API, and that's so normalized that people hardly blink anymore Anthropic inserts some things in your LLM context even when accessing it via API, impossible to avoid OpenAI never send you the raw reasoning, has a compaction endpoint that gives encrypted stuff, and even to access summaries you have to present "papers please". But I see your point. It feels like a tide of increasing closeness

> Also wondering what is keeping DeepSeek from making a new release. Their last release is from Dec 2025. as i heard, they test and try china's hardware to avoid sanctions issue; they almost did the next ver of llm - it was on lmarena (for some time) with the same name as prev ver but better results; my colleagues noticed this change

Yeah, I’m concerned about this too. I’ve found Nathan Lambert’s X posts about this informative and also a little bit depressing: https://x.com/natolambert His take is that basically given the capital expenditure involved, companies that are making LLMs as a business will likely have trouble if they make purely open models. OTOH, there might be some companies like NVIDIA (or Meta?) that have other core businesses that could afford to do it.

I actually think the "talk to us for commercial use" model MiniMax is exploring can be a good model for financial sustainability. It's only fair that inference providers like Fireworks work out a licensing deal to serve their models, and they're happy to do so with day 0 support. MiniMax has clarified that "personal" self-hosting even for writing commercial software projects is completely allowed. Qwen 3.6 released weights for the 35B-A3B model today, Apache licensed. That was a great relief and good sign of things to come. Also great to see its great performance, edging out the similarly-sized Gemma 4. I wrote a bit about it here: https://aicoe.fit/qwen-36-open-vs-opus-47-vs-gemma-7dcba5

Nice article! Re: MiniMax license: MiniMax has been going back and forth with their license in recent times. It started with the Dpt of War and Anthropic incident, which prompted them to add clauses to prevent military use and citizen abuse. They then moved to a non-commercial license for MiniMax-M2.7. They later patched it to allow “personal” use but the license broadly remains non-commercial. So, as it stands, it disallows to exploit the model commercially, even for personal projects. There are colorful threads in HF discussion forums about this. More broadly, no enterprise would adopt such a patched license and they’d need to agree on a cleaner license. If only MiniMax does this, I guess they exclude themselves from enterprise. If others follow suit, the landscape changes dramatically.