Most "best open-source LLM" rankings are quality rankings — which model gets the highest benchmark score. That's not what "open source" means. A model with published weights under a license that forbids commercial use is not open source in the sense that, say, Linux or PostgreSQL is. This post ranks the 14 best genuinely-usable LLMs in 2026 by license freedom first, then quality. The top of the list is models you can use however you want. The bottom is "open weights" models with license strings attached.
Short version: Qwen 2.5 is the best truly-open option right now. Llama 3.2 and 3.3 are excellent but not technically open source under the OSI definition. Gemma 2 is somewhere in between. Scroll down for the full ranking, or see our runtime-focused roundup for pure quality picks without the licensing lens.
How "open source" actually breaks down in 2026
The Open Source Initiative published a formal definition of open-source AI in late 2024, and most of the "open" models don't technically qualify. The big dividing lines:
- Weights published under a permissive license (Apache 2.0, MIT): unambiguously open source. You can commercialize, fine-tune, redistribute.
- Weights published under a custom non-commercial or conditional license (Llama Community, Gemma Terms, DeepSeek): often called "open weights" or "source-available." Free for most use but with legally binding conditions.
- Weights under non-commercial research licenses (Mistral NC, StableLM NC): free for personal and academic use only.
- Training data, code, and weights all published: the OSI's actual bar. Only a handful of models meet it (OLMo, Pythia, some smaller research models).
How we ranked
- License freedom (40%): OSI-approved > Apache/MIT > custom permissive > non-commercial.
- Quality (30%): Average across MMLU, HumanEval, and GSM8K from primary sources.
- Ecosystem (15%): Runtime support (llama.cpp, MLX, Ollama, vLLM) and fine-tune availability.
- Transparency (10%): Training data disclosure, model cards, reproducibility.
- Practicality (5%): Size relative to hardware availability.
The 14 best open-source LLMs in 2026
1. Qwen 2.5 7B — 94/100 (Apache 2.0)
The best truly-open model you can actually run. Apache 2.0, benchmark-competitive with Llama 3.1 8B and beating it on most tasks, 4.5 GB at Q4. Alibaba's decision to release the 1.5B through 32B variants under Apache 2.0 (the 72B has a different license) is the biggest open-weights story of the year. If your organization cares about license cleanliness, this is your baseline.
2. Phi-3.5 Mini 3.8B — 92/100 (MIT)
Microsoft's small model under MIT license. MIT is as permissive as it gets. Quality is exceptional for its size, particularly on reasoning and code. 2.4 GB at Q4, runs on phones. If you want the smallest possible truly-open model that's still good, this is it.
3. Qwen 2.5 Coder 32B — 90/100 (Apache 2.0)
The best open-weights coding model. Apache 2.0. HumanEval in the high 80s. Needs serious hardware to run (~20 GB at Q4), but for organizations with workstation budget, this replaces paid code-assist tools with a fully-controllable local model. Covered in depth in our best coding LLMs post.
4. Mistral Nemo 12B — 87/100 (Apache 2.0)
Mistral + NVIDIA's joint release. Apache 2.0. 128K context window. One of the few open models designed from day one to be quantization-friendly. The best "I have a 16 GB laptop and want the best open model that fits" choice.
5. Qwen 2.5 Coder 1.5B — 85/100 (Apache 2.0)
The best sub-2B open coding model. Apache 2.0. Small enough to run on a phone, HumanEval in the low 60s. Use as a local completion engine or lightweight code-assist.
6. SmolLM2 1.7B — 83/100 (Apache 2.0)
Hugging Face's in-house small model. Apache 2.0. Trained on a carefully filtered, documented dataset — the training recipe is the most transparent of any top-10 model on this list. Small, fast, genuinely open in the full sense.
7. OLMo 2 7B — 80/100 (Apache 2.0)
Allen Institute for AI's OLMo 2 is one of the only models that meets the full OSI definition: weights, training code, training data, and evaluation harness are all published. Quality is slightly behind Qwen 2.5 7B, but if reproducibility and scientific transparency matter to you, OLMo 2 is the only serious option at this size.
8. Llama 3.2 3B — 78/100 (Llama Community)
Meta's 3B is arguably the best 3B model on pure quality, but it ships under the Llama Community License, not Apache or MIT. The license is permissive for most users (free for commercial use up to 700M monthly active users), but it is not OSI-approved. Significantly weighted down here on license, but the raw quality is hard to beat at this size.
9. Llama 3.1 8B — 77/100 (Llama Community)
Same story: excellent quality, non-OSI license. The workhorse of the current Llama generation and the model with the largest fine-tuning ecosystem. Cleaner than Llama 3.2 in license terms (same license, but released earlier and more thoroughly battle-tested in legal review).
10. Gemma 2 2B — 75/100 (Gemma Terms)
Google's 2B small model. Excellent quality for its size, strong multilingual performance, clean safety tuning. Gemma Terms are more permissive than Llama's (no MAU cap) but still not OSI-compliant. A good second-tier-open choice.
11. Llama 3.3 70B — 72/100 (Llama Community)
Meta's 70B "compact" model — much cheaper to serve than the older 405B while retaining most of the quality. Same Llama license caveats. You need workstation hardware (48 GB+ RAM at Q4) to run it locally.
12. DeepSeek V3 — 70/100 (DeepSeek License)
DeepSeek's latest flagship. Benchmark-competitive with the frontier at a fraction of the cost to train and serve. The DeepSeek license is custom and not OSI-approved, with some conditions that require careful reading for commercial deployment. Quality is excellent; license cleanliness less so.
13. Mistral 7B v0.3 — 68/100 (Apache 2.0)
The original open Mistral. Apache 2.0. Now comfortably beaten on quality by Qwen 2.5 7B and Llama 3.1 8B, but still an excellent model with an enormous fine-tune ecosystem. Include here because of license cleanliness combined with maturity.
14. Pythia 12B — 60/100 (Apache 2.0)
EleutherAI's Pythia is one of the earliest fully-transparent LLMs — training data, code, checkpoints throughout training, all under Apache 2.0. Quality is dated in 2026, but Pythia remains the benchmark for scientific reproducibility. Include it if research reproducibility is your top priority.
The comparison table
| # | Model | License | OSI-open? | Commercial OK? | Score |
|---|---|---|---|---|---|
| 1 | Qwen 2.5 7B | Apache 2.0 | Yes | Yes | 94 |
| 2 | Phi-3.5 Mini | MIT | Yes | Yes | 92 |
| 3 | Qwen 2.5 Coder 32B | Apache 2.0 | Yes | Yes | 90 |
| 4 | Mistral Nemo 12B | Apache 2.0 | Yes | Yes | 87 |
| 5 | Qwen 2.5 Coder 1.5B | Apache 2.0 | Yes | Yes | 85 |
| 6 | SmolLM2 1.7B | Apache 2.0 | Yes | Yes | 83 |
| 7 | OLMo 2 7B | Apache 2.0 | Yes (full) | Yes | 80 |
| 8 | Llama 3.2 3B | Llama Community | No | Conditional | 78 |
| 9 | Llama 3.1 8B | Llama Community | No | Conditional | 77 |
| 10 | Gemma 2 2B | Gemma Terms | No | Yes (restricted) | 75 |
| 11 | Llama 3.3 70B | Llama Community | No | Conditional | 72 |
| 12 | DeepSeek V3 | DeepSeek | No | Yes (read license) | 70 |
| 13 | Mistral 7B v0.3 | Apache 2.0 | Yes | Yes | 68 |
| 14 | Pythia 12B | Apache 2.0 | Yes (full) | Yes | 60 |
Which open-source LLM should you pick?
For commercial products: Qwen 2.5 7B. Apache 2.0, competitive quality, no MAU caps, no surprises in the license. This is the right default for any business.
For on-device deployment: Phi-3.5 Mini (MIT) or Qwen 2.5 Coder 1.5B (Apache 2.0). Both are legitimately open and both fit on a phone. PocketLLM ships these as one-tap downloads — join the waitlist.
For research reproducibility: OLMo 2 or Pythia. They're the only two models on this list where you can inspect the training data and reproduce the training run.
For "best quality, don't care about license strictness": Llama 3.3 70B if you have the hardware, Qwen 2.5 32B or Llama 3.1 8B otherwise. See our general-purpose LLM ranking if license isn't your primary lens.
The quick answer
In 2026, the best genuinely open-source LLM is Qwen 2.5 7B. The best openly-licensed small model is Phi-3.5 Mini. The best reproducible research model is OLMo 2. Everything else called "open" is some flavor of source-available — still useful, but read the license before you commercialize.