Llama 4, Mistral Large 3, DeepSeek V3, and Qwen 3 have closed the gap with proprietary models.
The State of Open Source LLMs In the past 12 months, the open source LLM landscape has transformed dramatically. Where once proprietary models held a clear lead in capability, the gap has narrowed to a sliver — and for many use cases, closed entirely. The Big Four Llama 4 (Meta) ships with 8B, 70B, and 400B variants, all under a commercial friendly license. The 1M token context window matches Gemini's. Mistral Large 3 brings native multilingual reasoning — the first open model with parity in Arabic and Mandarin benchmarks vs. English. DeepSeek V3 delivers GPT 4 class reasoning at 1/30th the inference cost. Qwen 3 from Alibaba dominates Chinese language tasks and offers strong multimodal capability. What This Means For builders: you can now self host capable models. For incumbents: pricing pressure is real.