Qwen 3
Alibaba’s May 2025 release — first open family with hybrid reasoning (toggle-able chain of thought), Apache 2.0 across all sizes.
Benchmark: MT-Bench
Llama 4 (Scout/Maverick)
Meta’s April 2025 MoE-and-multimodal release, headlined by Scout’s 10M-token context window and the pre-announced Behemoth frontier model.
Gemma 3
Google DeepMind’s March 2025 Gemma family — vision-capable (4B+), 128K context, with official quantization-aware 4-bit variants.
Mistral Small 3
Mistral AI’s January 2025 24B model — Apache 2.0, competitive with Llama 3.3 70B, fits on a single 24GB GPU.
DeepSeek R1
DeepSeek’s January 2025 reasoning model — frontier chain-of-thought quality, plus six MIT-licensed distills from 1.5B to 70B.
DeepSeek V3
DeepSeek’s December 2024 frontier-scale MoE — 671B total, 37B active, trained for ~$5.6M in compute.
Phi-4
Microsoft Research’s December 2024 Phi-4 — a 14B dense MIT-licensed model punching well above its weight on math and reasoning.
Llama 3.3
A single 70B model released December 2024, closing most of the gap to Llama 3.1 405B through improved post-training alone.
Llama 3.2
Meta’s September 2024 Llama release added edge sizes (1B/3B) and the first open-weight Llama vision models (11B/90B).
Qwen 2.5
Alibaba’s September 2024 Qwen family spans 0.5B to 72B, plus coding and math specialists — mostly Apache 2.0.
Llama 3.1
Meta’s flagship 2024 open-weight LLM family — 8B, 70B, and 405B parameters with 128K context. The 405B was the first open-weight model at true frontier scale.
Gemma 2
Google DeepMind’s June 2024 lightweight open model family — 2B, 9B, and 27B with interleaved local/global attention.
Mixtral 8x7B
Mistral AI’s December 2023 mixture-of-experts model — 8 experts, 2 active per token, Apache 2.0, ran at Llama-13B speed with Llama-70B quality.
Mistral 7B
Mistral AI’s September 2023 debut — a 7B Apache-2.0 model that popularized Grouped-Query and Sliding Window Attention.