LLM·Dex
Use caseTop 8 picks

Best Open-Source LLMs in 2026

The top open-weight models for self-hosting or fine-tuning.

Updated

How we ranked

  • Composite benchmark performance
  • License permissiveness (Apache, MIT, custom OSS)
  • Inference economics on commodity GPUs
  • Fine-tuning ecosystem maturity
  • Multilingual coverage

Read the full methodology for our sourcing and ranking standards.

Open-weight models are no longer "the cheap alternative", they're competitive. Llama-4-405B trades blows with GPT-5 on standard benchmarks. DeepSeek-V3 and R1 outperform every closed model in cost-per-quality terms. Qwen-3 leads on Chinese and matches the West on English.

For self-hosters, the right choice depends on your hardware and your license tolerance. Llama uses a custom community license; Apache-2 lovers should look at Mistral, Qwen, and DeepSeek. For fine-tuning, the Llama and Qwen ecosystems have the most mature tooling.

Don't conflate open-weight with open-source. Most "open" model releases ship the weights but not the training data, and the training-data gap is where most reproducibility ends.

The ranking

  1. #1MetaOpen weights

    Llama 4 405B

    Meta's flagship open-weight model, sparse MoE design competitive with closed-frontier flagships.

    Context
    256K tokens
    Output · 1M
    Pricing not published
    Modalities
    text, vision

    Why it ranks here. Strongest open-weight model at launch. Multimodal (text + vision). Tracked weakness: Custom license with revenue caps.

  2. #2DeepSeekOpen weights

    DeepSeek-V3

    DeepSeek's flagship 671B-parameter MoE, frontier-level quality at a tiny fraction of frontier prices.

    Context
    128K tokens
    Output · 1M
    $1.10 / 1M tokens
    Modalities
    text

    Why it ranks here. Frontier-level quality at open-weight prices. MIT license, clean commercial use. Tracked weakness: No native vision support.

  3. #3AlibabaOpen weights

    Qwen3-72B

    Alibaba's flagship open-weight Qwen3, strong on multilingual, code, and math, Apache-2.0 licensed.

    Context
    128K tokens
    Output · 1M
    Pricing not published
    Modalities
    text

    Why it ranks here. Apache-2.0 license. Strongest open-weight on Chinese. Tracked weakness: No native vision in this variant.

  4. #4MetaOpen weights

    Llama 4 70B

    Meta's mid-tier Llama 4, the practical workhorse for self-hosted deployments.

    Context
    128K tokens
    Output · 1M
    Pricing not published
    Modalities
    text, vision

    Why it ranks here. Self-hostable on commodity hardware. Strong all-rounder. Tracked weakness: Custom license.

  5. #5DeepSeekOpen weights

    DeepSeek-R1

    First open-weight reasoning model to match o1, the release that proved RL-from-scratch reasoning training was reproducible.

    Context
    128K tokens
    Output · 1M
    $2.19 / 1M tokens
    Modalities
    text

    Why it ranks here. Open-weight reasoning model on par with o1. MIT license. Tracked weakness: Slow, reasoning is slow by design.

  6. #6MistralOpen weights

    Mixtral 8×22B

    Mistral's largest open-weight MoE, Apache-2.0, still widely deployed.

    Context
    64K tokens
    Output · 1M
    Pricing not published
    Modalities
    text

    Why it ranks here. Apache-2.0. MoE economics. Tracked weakness: Older generation.

  7. #7AlibabaOpen weights

    Qwen2.5-72B

    The previous-generation Qwen flagship, still widely deployed for stability.

    Context
    128K tokens
    Output · 1M
    Pricing not published
    Modalities
    text

    Why it ranks here. Mature deployment. Apache-2.0. Tracked weakness: Superseded by Qwen3 for new builds.

  8. #8MetaOpen weights

    Llama 3.3 70B

    Meta's late-2024 70B refresh, much-improved over 3.1 with better instruction-following and tool-use.

    Context
    128K tokens
    Output · 1M
    Pricing not published
    Modalities
    text

    Why it ranks here. Mature, well-tuned. 70B fits on a single H100. Tracked weakness: No native vision.

How to choose

Don't pick on the headline ranking alone. Run your top two picks on a representative sample of your own workload and compare. The numbers in this list are sound, but task-specific quality varies in ways no benchmark fully captures. The criteria above are the right axes to evaluate on, but the weighting depends on your stack.

  • Cost-sensitive workloads, start with the cheapest of the top three; escalate only if quality is the bottleneck.
  • Privacy-sensitive workloads, filter to open-weight picks above. They're labeled with a green badge.
  • Latency-sensitive workloads, see the Fastest LLMs list, which can override task-specific picks.

Frequently asked

  • What is the best model for open-source llms?
    Our #1 pick is Llama 4 405B from Meta. Meta's flagship open-weight model, sparse MoE design competitive with closed-frontier flagships.
  • How are these rankings determined?
    We rank by the criteria listed at the top of this page: Composite benchmark performance; License permissiveness (Apache, MIT, custom OSS); Inference economics on commodity GPUs. Where two models are close, we prefer the one with stronger production deployment evidence at the time of writing. Read the full methodology for our standards.
  • Llama 4 405B or DeepSeek-V3?
    Both are top-tier picks. Llama 4 405B edges ahead on the criteria most relevant to this task. DeepSeek-V3 is the strongest alternative, see the head-to-head comparison page for full deltas.
  • Are open-source models on this list?
    Yes where they're competitive. Each entry below shows whether the model ships open weights and under what license.
  • How often is this list updated?
    Weekly. New launches that affect the ranking get reflected within seven days. The "last updated" stamp at the top of the page reflects the most recent dataset commit.

Related guides

Friday digest

The week's AI launches, in your inbox.

One short email every Friday, new models, leaks, and quietly-shipped APIs you missed.