Open Source LLM Comparison (2025–2026)

Meta's fourth generation, released in April 2025. The Llama 4 Scout (17B active, 109B total MoE) and Maverick (17B active, 400B total) are multimodal, trained on 10M+ image-text pairs. Maverick matches GPT-4o on several benchmarks. Scout fits on a single H100 GPU and handles 10M token context via iRoPE position encoding.

Licence: Meta Llama 4 Community Licence - free for commercial use with fewer than 700M MAU

Mistral 7B (Apache 2.0, genuinely open) remains the most popular small model for local deployment. Mistral Small and Mistral Large are available via API; weights released under a Mistral Research Licence. Codestral is a code-specialised variant. Mixtral 8x7B and 8x22B are strong mixture-of-experts models for inference at lower effective parameter cost.

Best for: European data residency requirements (French company), code generation (Codestral)

Gemma 3 (released March 2025) comes in 1B, 4B, 12B, and 27B parameter sizes. Strong instruction following, multilingual (140+ languages), and optimised for edge deployment on Google hardware. The 27B model outperforms Llama 3.1 70B on several benchmarks. Available with quantized variants for CPU inference.

Best for: edge deployment, Google Cloud integration, multilingual applications

Qwen 2.5 is available in 0.5B to 72B parameters. Outstanding multilingual performance (strongest for CJK languages among open models). Qwen 2.5-Coder is a leading open code model. The 72B instruction-tuned model matches or exceeds Llama 3.1 70B on most benchmarks. Most sizes are Apache 2.0; 72B has a custom licence.

Best for: Chinese language applications, coding tasks (Coder variant), cost-efficiency

DeepSeek V3 (released December 2024) shocked the industry by matching GPT-4o on coding and maths benchmarks at a fraction of the training cost. DeepSeek R1 (January 2025) is a reasoning model with visible chain-of-thought, matching o1 on many tasks. Both are MIT licenced - the most permissive licence of any frontier-class model.

Best for: cost-sensitive deployments, reasoning tasks; note: China-based company, data considerations for regulated industries

Microsoft's Phi series demonstrates that quality training data can compensate for fewer parameters. Phi-4 (14B) outperforms several larger models on reasoning and coding tasks. Excellent for edge deployment where parameter efficiency matters.

Best for: edge/mobile deployment, reasoning tasks where parameter efficiency matters