Open-Source LLM Competition Heats Up: Llama 4, Mistral Large 2 and More

The open-source large model space is seeing unprecedented competition, with new-generation models like Llama 4 and Mistral Large 2 launching in quick succession.

These open models keep closing in on — and in some tasks matching — closed flagships, while offering the advantages of self-hosting, fine-tuning and cost control, attracting many enterprises and developers.

The thriving open-source ecosystem also drives rapid iteration of toolchains, fine-tuning methods and inference optimization. Analysts say open and closed models will coexist and compete for the long term, together accelerating the industry's technical progress and adoption.

The Center of Gravity Shifts East

After 2025 the open-source story is no longer defined by Llama: DeepSeek R1 detonated open reasoning, Qwen3 covered every size class, and Kimi K2 topped the charts at a trillion parameters (see our coverage). Llama 4's lukewarm reception and Behemoth's stall cost it the 'synonym for open source' spot. The center of gravity in open-source competition is moving from the U.S. to the Chinese relay.

From Capability Parity to Price-Performance Dominance

Open models not only close in on closed ones at the benchmark level but push deployment cost to a fraction of comparable performance: self-hostable, fine-tunable, data never leaving the building — traits that hit the core demand of enterprise self-hosting (see our local deployment review). Even OpenAI returned to open weights with gpt-oss, proving open source has shifted from an ideological fight to a gateway fight.

Our Take

Open versus closed is not zero-sum: closed flagships raise the capability ceiling, open relays lower the barrier to entry, together lifting the industry floor. For developers the real dividend is choice: self-host open small models for light daily tasks, call closed-flagship APIs for heavy ones — the hybrid architecture is becoming default practice.