Appearance
The moat fell last month
Last month, the open source LLM moat crossed.
Not incrementally. Not "closing the gap". Four separate teams shipped open weight models that beat closed frontier models on their strongest benchmarks. None of them are research previews. All are production ready, deployed, and available to run on your hardware today.
This is not a normal release cycle. This is the shift Mark Zuckerberg wrote about when he released Llama 3.1. The Unix to Linux moment for AI arrived faster than almost anyone predicted.
Llama 3.1 405B: The standard was set
Meta did not just release another model. They declared open source as the end state for the industry.
Llama 3.1 405B is the first frontier model released with full weights. It runs on your own infrastructure at roughly 50% the cost per token of GPT-4o. It matches closed models on almost all general reasoning benchmarks, and beats them on fine tuning suitability.
This release was never just about the model. Meta lined up every major cloud, every inference provider, every enterprise services firm before launch. AWS, Databricks, NVIDIA, Groq, Scale all had production support available on day zero. This was not a model drop. This was an ecosystem declaration.
Meta has one structural advantage no closed provider can match. They do not make money selling model access. Releasing Llama does not cannibalize their revenue. It destroys their competitors' revenue. Every organization that moves to open source models is one less customer for OpenAI, Anthropic and Google. This incentive mismatch will accelerate every release from this point forward.
MiniMax M3: No one saw this coming
MiniMax M3 caught every production ML team completely off guard.
This is the first open model that beats closed frontier models on agent and coding benchmarks. It scores 83.5 on BrowseComp, ahead of Claude Opus 4.7 at 79.3. On PostTrainBench it ranks third globally, behind only GPT-5.5 and Opus, and ahead of every other model open or closed.
M3 has native 1M token context, built on a custom sparse attention architecture. It has true end to end multimodal alignment trained from initialization, not bolted on after pretraining.
The demonstrations are not cherry picked leaderboard scores. The team tasked M3 with reproducing an ICLR outstanding paper unassisted. It ran for 12 hours, produced 18 commits and 23 experimental figures, and successfully replicated all core results. When asked to optimize an FP8 GEMM kernel for Hopper GPUs it ran 147 iterations over 24 hours and delivered a 9.4x speedup with zero human intervention.
Weights will be published on Hugging Face. No one outside MiniMax saw this coming. Most teams still have not processed that this capability now exists in open weights.
Mellum2: The quiet most useful release this quarter
JetBrains did not make any noise. Nobody is writing hot takes about Mellum2. Every production ML team is already rolling it out.
Mellum2 is a 12B parameter Mixture of Experts model that activates only 2.5B parameters per token. It delivers benchmark performance comparable to 70B dense models while running more than 2x faster. It is released under full Apache 2.0 license with no restrictions.
This is not a frontier model. It is the workhorse model.
For three years every team has been building custom distilled models for routing, RAG compression, prompt classification, tool selection and intermediate agent steps. JetBrains just shipped a better one, free, no strings attached.
Mellum2 will be running 70% of all intermediate model calls in production systems by the end of this year. It will never top a public leaderboard. It will save more engineering hours and more inference cost than every other model released this quarter combined.
Cosmos 3: Physical AI stops being vaporware
NVIDIA shipped the most important model that no one is talking about for general LLM discourse.
Cosmos 3 is the first open foundation model built for the physical world. Before this release every robotics and autonomous vehicle team maintained four separate models for perception, scene reasoning, forward prediction and policy generation. Cosmos 3 does all four in a single unified forward pass.
It uses a custom Mixture of Transformers architecture that runs autoregressive reasoning and diffusion generation in the same layer with joint attention. It accepts text, images, video, audio and action inputs. It outputs any combination of those same modalities.
Two variants are available. The 16B Nano model runs at interactive speeds on a single workstation RTX PRO 6000 GPU. The 64B Super variant runs on Hopper and Blackwell for large scale synthetic data generation.
This model will not appear on any standard LLM benchmark. It will change how every physical AI system is built over the next 18 months.
Nobody is competing on trivia benchmarks anymore
For three years every model release followed exactly the same script. Vendors would announce a new model, post a 2 point improvement on MMLU or GSM8K, and everyone would argue for three days about benchmark contamination.
That era is over. None of the four models released this cycle lead general knowledge benchmarks. None of them even tried.
Every team is now competing on capabilities that actually matter for production:
- Long context that actually retains information across 100k+ tokens
- Agent execution that runs for hours without deviating from the task
- Inference throughput and latency
- Deployability on commodity hardware
- License terms that do not lock you into a single vendor
Nobody cares if your model can win trivia anymore. They care if it can optimize a CUDA kernel unassisted. They care if it can run 1000 routing requests per second on one GPU. They care if it can simulate a robot picking up an object.
The new production stack everyone is building right now
Right now every senior ML engineering team is rebuilding their production stack to exactly this pattern:
Frontier reasoning and hard coding tasks run on either Llama 3.1 405B or MiniMax M3. All intermediate work runs on Mellum2. Any physical world, robotics or simulation work runs on Cosmos 3 Nano.
That is the entire stack.
No more GPT-4o API calls for 90% of workloads. No more waiting for closed vendor roadmaps. No more rate limits. No more terms of service changes. You can run all three models on your own hardware today, fine tune them, modify them, and never have to ask permission.
This transition is happening quietly. Most public commentary has not caught up yet. Teams are not tweeting about it. They are just migrating workloads.
The closed model advantage is now just brand
GPT-5.5 and Claude Opus are still slightly better at some very hard open ended reasoning tasks. The gap is now less than six months. It is closing faster with every release.
Closed providers now have exactly one remaining advantage: marketing. Most non technical decision makers still have not internalized that open models are now good enough for almost all production use cases. That perception will shift this year.
Closed providers will continue to lobby governments for regulation that restricts open source models. They have no other moat left.
What comes next
Meta stated publicly that they expect Llama 4 to be the best model in the world next year. That is no longer an empty claim. Every release now validates that open development is accelerating faster than closed development.
The next step is not bigger models. It is better interfaces between specialized models. It is standard tooling that works the same way across every model. It is production reliability.
We are no longer waiting for the frontier to arrive. It is here. It is open. You can download it tonight.
References
- Meta, Open source AI is the path forward: https://about.fb.com/news/2024/07/open-source-ai-is-the-path-forward/
- MiniMax M3 Model Page: https://www.minimax.io/models/text/m3
- JetBrains Mellum 2 Launch: https://huggingface.co/blog/JetBrains/mellum2-launch
- NVIDIA Cosmos 3 Launch: https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai
- StepFun Step 3.7 Flash: https://huggingface.co/models/stepfun-ai/Step-3.7-Flash