Ai

Open Source LLM Review 2026: 2.8T MoE and MiniMax M3 Arrive

Open Source LLM Review 2026: 2.8T MoE and MiniMax M3 Arrive

The open-source LLM race is showing no signs of slowing. Two new releases in June and July 2026 have pushed the boundaries of what's possible with open weights: a 2.8-trillion-parameter sparse mixture-of-experts model now available on Fireworks, and MiniMax's M3, a 428B-parameter model that's already turning heads with its unusual licensing terms. Here's what changed and why it matters.

The 2.8T Sparse MoE: Massive Scale, Sparse Activation

The first model is a beast. With 2.8 trillion total parameters and 896 experts, it's one of the largest open-weight models ever built. But here's the trick: it only activates 16 experts for each token. That's roughly 1.8% of the experts, which sounds tiny. But it's enough to deliver strong performance on hard benchmarks. The model scores 93.5% on GPQA, a graduate-level reasoning test, and 76.2 on the Artificial Analysis Coding Index. Those numbers put it in the same conversation as some of the best closed models.

Then there's the pricing. At $3 per million input tokens and $15 per million output tokens, it's not the cheapest model out there. But for a model this size, the price feels almost too good. The 1 million token context window — that's about 850 pages of text — makes it a natural fit for document-heavy tasks. Legal review, long-form code analysis, you name it.

The full weights are set to drop on July 27. That's the real story. Once the weights are out, the community can fine-tune it, run it in production, and build on top of it. Right now, you can only access it through Fireworks, but that's about to change.

MiniMax M3: A Practical Contender with a Quirky License

MiniMax M3 takes a different path. It's a 428B-parameter model with only 23B activated per token. That means it's not just a research curiosity — it's something you can actually run on a single node, or at least a modest cluster, without breaking the bank.

The context window is 512k tokens on Fireworks. Not as big as the 2.8T model, but plenty for most enterprise use cases. And the performance? MiniMax hasn't published a full benchmark suite, but the model has been in the wild for about a month, and early reports have been solid.

What's most interesting, though, is the license. The MiniMax Community license requires commercial deployments to display 'Built with MiniMax M3.' If you're under $20 million in annual revenue, a one-time notice is all you need. Above that, you need prior written authorization. That's a two-tiered licensing structure we haven't seen before from a major AI lab. It's a way to hook smaller startups early, then negotiate bigger deals once they scale.

Why does this matter? Because it shows that open source doesn't have to mean free-for-all. MiniMax is building a business model around controlled openness, and it could be a blueprint for other labs trying to balance community goodwill with commercial viability.

What This Means for the Open Source Landscape

We're seeing a clear split in 2026. Some labs are racing to push parameter counts higher and higher, betting on scale as the path to intelligence. Others are focusing on efficiency and practical deployment, with licensing that lets developers ship today without legal hand-wringing. Both approaches have merit.

The 2.8T model is an impressive engineering feat, but its day-to-day utility depends on whether the full release delivers on its promise. MiniMax M3, meanwhile, is already usable for production apps, and its license gives developers a predictable framework for commercial use.

In my view, the biggest takeaway isn't the benchmark scores. It's the speed of the release cycle. These models went from announcement to API availability in a matter of weeks. The open source ecosystem is moving at a pace that's hard to keep up with. That's a good problem to have.

Official Source: https://fireworks.ai/blog/best-open-source-llms