Open Weight Models — May 2026 Landscape

The open-weight LLM landscape in May 2026 is genuinely competitive with closed APIs for most tasks. Llama 4 (Meta, April 2025), Qwen 3 (Alibaba), DeepSeek V3 / R1, Mistral, Phi-4 (Microsoft), Gemma 3 (Google), and Command-R+ (Cohere) are the names you should know. This page is the map; specialized lanes covered separately: coding models, vision / multimodal, embeddings & rerankers, quantization.

For "what runs on my hardware," see the hardware tier guide. For inference engines see Ollama, llama.cpp, vLLM / SGLang.

The frontier-class open families

Llama 4 (Meta, April 2025)

• ★ ★ Released in multiple sizes including a flagship MoE (~400B+ total, ~17B active per token) and smaller dense variants.
• License: Llama 4 Community License — commercially permissive with the >700M-MAU clause Meta has used since Llama 2. Read it.
• Strong general capability; tool use; long context (1M+ on top variants); multimodal in some variants.
• The dominant frontier-class open model in 2026.

Qwen 3 (Alibaba, late 2024 / 2025)

• ★ ★ Multiple sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B; plus MoE variants.
• License: Apache 2.0 for most sizes — extremely permissive.
• Excellent multilingual; particularly strong in Chinese / Asian languages.
• Strong tool-use, structured output, coding (paired with Qwen 2.5/3 Coder).
• ★ The "Apache 2.0 frontier" pick — best license + capability tradeoff.

DeepSeek V3 (general) and R1 (reasoning)

• ★ ★ DeepSeek V3 — large MoE (~671B total, ~37B active per token); SOTA-class general capability when it released late 2024.
• ★ ★ DeepSeek R1 — reasoning-focused; "show your work" thinking traces; strong math / code.
• License: DeepSeek License (custom, broadly permissive for most uses; check terms for derivative training).
• R1 distillations into Llama / Qwen 3 base sizes give you reasoning-grade models that fit on consumer hardware.
• ★ The home-server reasoning standard, particularly the R1-distill family.

Mistral / Mixtral / Codestral / Mistral Small 3 / Large 2

• ★ Mistral 7B started the open-weight wave; Mistral keeps shipping.
• ★ Mistral Large 2 — 123B dense; strong general capability.
• ★ Mixtral family — earlier MoE pioneers (8x7B, 8x22B); still useful.
• ★ Codestral — coding-specialist.
• ★ Mistral Small 3 — efficient general-purpose.
• License: mostly Apache 2.0 for smaller models; Mistral Research License (non-commercial only) for the larger ones — read carefully.

Phi-4 (Microsoft)

• ★ Phi-4 14B and Phi-4-mini — punches well above its weight on reasoning / math.
• License: MIT.
• Synthetic-data-trained; less general knowledge than Llama / Qwen at the same size; better at math / logic.
• ★ Best 14B-class for reasoning if you don't need broad knowledge.

Gemma 3 (Google)

• ★ Gemma 3 1B / 4B / 12B / 27B.
• License: Gemma Terms of Use — commercially usable with restrictions; not full Apache.
• Strong long-context; efficient memory use.
• Multimodal variants (Gemma 3 4B/12B vision).

Command-R+ (Cohere)

• ★ Command-R+ 2025 — RAG and tool-use specialist; extensively documented for those use cases.
• License: CC-BY-NC-4.0 for the open weights (non-commercial only); commercial use requires Cohere license.
• Excellent at structured outputs and tool routing.

What's not on this list (and why)

• GPT-5, Claude Opus 4.7, Gemini 2.5 Ultra — closed APIs, paid only. Honest framing: still meaningfully better than the best open weights at the very hardest reasoning, but the gap has narrowed to roughly 6–12 months. For workflows where they matter, see Aider with API or Claude Code.
• Grok-3 weights — xAI released some Grok weights in 2024 under a custom license; mostly a curiosity at the size that ran.
• Falcon, MPT, RedPajama — the 2023 wave; still around, mostly superseded.

Choosing a model for your hardware

License cheat sheet

• ★ ★ Apache 2.0: Qwen 3, Mistral 7B, Mixtral, Mistral Small 3, Phi-4, BGE embeddings.
• MIT: many Microsoft and community models.
• Llama 4 Community License: commercial OK below 700M MAU; otherwise contact Meta.
• Gemma Terms of Use: commercial OK with restrictions on harmful use.
• CC-BY-NC-4.0: Command-R+ — non-commercial only; commercial license from Cohere.
• DeepSeek License: custom, broadly permissive; read the latest version.
• Mistral Research License: non-commercial only on Mistral Large; smaller models are Apache.

How to read model names

• Llama-4-70B-Instruct — base 70 billion params, instruction-tuned.
• Qwen3-32B-Instruct — instruction-tuned chat variant.
• Llama-4-70B-Instruct-AWQ — AWQ-quantized version.
• Qwen2.5-Coder-32B — coding-specialist tune.
• DeepSeek-R1-Distill-Llama-70B — R1 reasoning distilled into a Llama 70B base.
• Quants: Q4_K_M, Q5_K_M, Q6_K, Q8_0, FP16 — see quantization.

Honest framing

• Benchmark games are real. Open-weight leaderboards (LMSYS Chatbot Arena, OpenRouter rankings, lmarena.ai) are useful but gameable. Check multiple sources.
• Day-1 capability you'll feel — instruction-following, math, coding, multilingual, hallucination rate.
• What benchmarks miss — vibes, tone, creative writing, in-character roleplay; the SillyTavern community has its own taste-led rankings.
• Re-evaluate every 6 months. The frontier moves fast; the model that was state-of-the-art when you set up your stack is not the one you should be running today.

Pick this if…

• One general-purpose pick, balanced license & capability: Qwen 3 32B (Apache 2.0).
• Best frontier-class at home: Llama 4 (largest variant your hardware fits).
• Reasoning-heavy workloads: DeepSeek R1 or its distillations.
• Coding-specialist: see coding models.
• Vision / multimodal: see vision models.
• Tiny models on tiny hardware: Phi-4-mini, Llama 3.2 1B/3B.
• Strict commercial license: stick to Apache 2.0 / MIT models.