If you'd asked us this in 2024, we'd have told you "Llama 3 and stop overthinking." In 2026 the answer is more interesting — DeepSeek owns the leaderboard, Phi-4 is freakishly efficient, and Gemma 3 ships small enough to run on your phone. Here's how we'd actually pick.
Frontier reasoning + open weights → DeepSeek V4
DeepSeek V4 is the first open-weights model that genuinely competes with GPT-5 and Claude Opus on hard reasoning benchmarks. The 685B MoE flagship needs an H100 cluster, but the 67B distilled variant lands on two A100 80GBs and still beats Llama 4 70B on MMLU. License is permissive — you can ship products on it.
When this is wrong: You don't have GPUs. The 67B distill is the smallest variant and you'll burn $$$ on a managed inference provider if you don't host yourself. For under-$5/M-token inference on a frontier-quality open model, see our LLM Pricing Calculator for hosted options.
Best 8B local model → Llama 4 8B (still)
Yes, Phi-4 14B beats it on benchmarks. Yes, Qwen3.6 7B has more languages. But the Llama 4 8B has the deepest fine-tune ecosystem, the most quantizations, the most integration libraries, and the most StackOverflow answers when something breaks. For "I want to run an LLM locally and not fight my tools," it's still the right pick.
Best small model that punches up → Phi-4 14B
Microsoft trained Phi-4 heavily on synthetic high-quality data and it shows: 95.2 on GSM8K, 84.8 on MMLU, all from a 14B base. That's better math than Llama 4 70B from a quarter the params. The catch: 16K context and English-heavy. For a math tutor, code helper, or anything where you can keep prompts short, it's the most efficient model on this list.
Best for fine-tuning your own model → Mistral 22B or OLMo 2
Two opposite reasons. Mistral 22B has the cleanest base model and a real research community around it. OLMo 2 from Allen AI publishes the full training code, data, and checkpoints — if you want to study why a model behaves a certain way, you can. Most production fine-tunes still happen on Llama, but the most reproducible research happens on these two.
Best multilingual → Qwen3.6 35B
119 languages out of the box, and the bilingual EN/CN performance is a level above what Llama or Mistral offer. If you're shipping to anywhere outside North America and Europe, start here. Vision is included on the 35B+ variants too.
Best for code → DeepSeek Coder V3
89.4 on HumanEval. Fill-in-the-middle support. Trained on 2T tokens of code. There's no close second. If your only use case is code generation, the 33B variant on a single A100 is the right answer.
What we did NOT recommend (and why)
- **Falcon 3 180B**: license tiers based on revenue, fewer fine-tunes, lags newer models on reasoning. Pick something else unless you specifically need Arabic.
- **Command R+ 2**: amazing for RAG but research-license-only weights. If you want it in production, you're paying for the Cohere API anyway.
- **Yi 1.5 34B**: solid but it's two years old now. Newer Apache-2.0 options (Mixtral 8x22B, OLMo 2) match it.
How to actually decide
Pick on these axes in order: **license** (research vs commercial), **VRAM you have** (8GB / 24GB / 80GB tiers), **single task or general** (code-only is much smaller), **languages you serve**. Most teams overthink the model and underthink the eval — pick a defensible default and run your own evals against your actual prompts.