Open-weight models can be downloaded, self-hosted, fine-tuned, and (with varying license restrictions) used commercially. They represent the best option for teams needing full control over their model deployment.
Meta Llama 3 / Llama 3.1 / Llama 4
Parameters Llama 3: 8B, 70B; Llama 3.1: 8B, 70B, 405B; Llama 4: Scout (17B active / 109B total MoE), Maverick (17B active / 400B total MoE) Context Length Llama 3.1: 128K; Llama 4: 10M tokens (Scout), 1M (Maverick) Architecture Decoder-only transformer; Llama 4 uses Mixture of Experts Key Strengths Strong baseline for fine-tuning, large ecosystem, excellent community support, competitive with proprietary models at scale License Llama Community License (commercial use allowed with <700M MAU restriction) Best For Custom fine-tuning, self-hosted inference, research, production deployments needing full control
Mistral AI: Mistral / Mixtral / Mistral Large
Parameters Mistral 7B; Mixtral 8x7B (47B total, 13B active), Mixtral 8x22B (141B total, 39B active); Mistral Large (123B); Mistral Small (24B) Context Length 32K (7B, Mixtral); 128K (Mistral Large) Architecture Decoder-only; Mixtral uses Sparse MoE with top-2 routing Key Strengths Excellent quality/size ratio, efficient MoE inference, strong multilingual support, function calling License Apache 2.0 (Mistral 7B, Mixtral); proprietary (Mistral Large) Best For Cost-effective self-hosting, multilingual applications, MoE experimentation
DeepSeek-V3 / DeepSeek-R1
Parameters DeepSeek-V3: 671B total (37B active MoE); DeepSeek-R1: 671B total (37B active MoE); R1 distilled variants: 1.5B, 7B, 8B, 14B, 32B, 70B Context Length 128K tokens Architecture MoE with Multi-Head Latent Attention (MLA) and auxiliary-loss-free load balancing Key Strengths V3: frontier-class general ability at low inference cost; R1: state-of-the-art open reasoning with visible chain-of-thought License MIT (fully permissive) Best For Cost-efficient self-hosting of frontier-quality models, reasoning tasks (R1), distilled reasoning (R1-distill variants)
Qwen 2.5 / QwQ
Parameters Qwen 2.5: 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B; QwQ-32B (reasoning variant) Context Length 32K (base); 128K (72B, QwQ) Architecture Decoder-only transformer with GQA, SwiGLU, RoPE Key Strengths Excellent multilingual ability (especially CJK languages), strong coding (Qwen2.5-Coder), competitive reasoning (QwQ) License Apache 2.0 (most sizes); Qwen License for 72B Best For Multilingual applications, coding assistants, fine-tuning base models, reasoning (QwQ)
Microsoft Phi-4 / Phi-4-mini
Parameters Phi-4: 14B; Phi-4-mini: 3.8B Context Length 16K tokens Architecture Decoder-only transformer; trained heavily on synthetic data and curated textbook-quality sources Key Strengths Exceptional performance at small sizes (especially math and reasoning), demonstrates the power of data quality over quantity License MIT Best For On-device deployment, edge inference, resource-constrained environments, research into data-efficient training
Google Gemma 3
Parameters 1B, 4B, 12B, 27B Context Length 32K tokens (text); 128K (27B) Modalities Text and image input (4B+); text output Key Strengths Strong multimodal understanding at small sizes, native vision capability, good fine-tuning base License Gemma Terms of Use (commercial use allowed with restrictions) Best For Multimodal applications on modest hardware, fine-tuning for specialized domains, on-device vision+language
License Differences Matter
Open weights does not mean open source. Check each model's license carefully: some (like Llama) restrict commercial use above certain revenue thresholds, while others (like Mistral's Apache-licensed models) have no such restrictions.
Proprietary Model Families
Model Cards and Selection Guide
Quick Comparison Table