The top 10 open-source LLMs

The short version

The interesting story in 2026 is not that open models exist. It is how close they now are. Epoch AI's tracking puts open-weight models roughly three months behind the state-of-the-art proprietary systems on average — a gap that was measured in years not long ago. For coding and reasoning in particular, the practical difference for most teams has narrowed to something you would struggle to notice day to day.

What has not narrowed is the licensing confusion, so we will deal with that first, because it changes which of these models you are actually allowed to use.

“Open weight” is not “open source”

Almost every model on this list is open weight, not open source. The distinction matters:

• Open source — weights, architecture, training code and (ideally) training data are public, under an OSI-recognised permissive licence such as MIT or Apache 2.0.
• Open weight — you can download and run the trained weights, but the training data and pipeline may be closed, and the licence may restrict commercial use, scale, or redistribution.

Most of the popular models sit in the second category. Meta's Llama family, for instance, ships under a bespoke community licence rather than a standard open-source one. If you are building a product on top of a model, read the licence before you read the benchmarks.

The ranking

Ordering leans on BenchLM's BenchAlign v5 composite index (snapshot 20 July 2026), cross-checked against other public leaderboards and task-specific results. Where sources disagree we say so.

1. MiniMax M3 — MiniMax

The current public open-weight leader, at 69.8 on the BenchAlign index with a 1M-token context. Released June 2026, it is reported as the first open-weight model to combine frontier-grade coding, million-token context and native multimodality in one package, and it tops the open-weight SWE-Bench Pro results at 59.0%. If you want the strongest general open model right now, this is it.

2. GLM-5.1 / GLM-5.2 — Z.AI

GLM-5.1 scores 67.7 with a 203K context and unusually broad benchmark coverage. The newer GLM-5.2 is described elsewhere as the strongest all-round open-weight model, and posts the highest GPQA Diamond score in its comparison set at 91.2%. Particularly well regarded for long-horizon coding agents — tasks that run for many steps without losing the plot.

3. Inkling — Thinking Machines Lab

A newer entrant sitting third at 67.5, also with a 1M-token context. Less established than the Chinese labs' releases, but it arrived near the top of the index immediately, which is worth watching.

4. DeepSeek V4 Pro — DeepSeek

The coding specialist. It leads SWE-bench at 80.6% and LiveCodeBench at 93.5%, with a 1M-token context, and DeepSeek's models remain the cost-effectiveness benchmark others are measured against. Its earlier reasoning model, DeepSeek R1 (671B MoE, 37B active), is still notable for shipping under a clean MIT licence.

5. Kimi K2.6 / K2.7 Code — Moonshot AI

A roughly 1.1T-parameter mixture-of-experts model under a modified MIT licence. The K2 line is the standout pick for agentic coding and UI generation — the kind of work where the model has to plan, call tools and iterate rather than autocomplete.

6. Qwen3.5-27B — Alibaba

Scores 60.7 with a 262K context. The Qwen family's real advantage is that the large variants (such as Qwen3 235B-A22B, 235B total / 22B active) ship under a genuinely permissive Apache 2.0 licence, and multilingual performance is among the best in open weights. For enterprises that need licence certainty, this is often the first stop.

7. Gemma 4 — Google

The practical laptop model. The 31B variant scores 61.1 with a 256K context, while the 26B A4B configuration (25.2B total, just 3.8B active) is built for local deployment on modest hardware. Apache 2.0. If you want something capable running on your own machine rather than someone's API, start here.

8. Llama 4 Scout — Meta

109B total / 17B active MoE, and the context outlier of the group at a claimed 10M tokens. The catch is the Llama Community Licence — widely used, but not OSI open source, with conditions that matter at scale. Still the default ecosystem: the tooling, fine-tunes and community support are unmatched.

9. Nemotron 3 Super — NVIDIA

The pick for teams who care about openness rather than just downloadable weights. NVIDIA publishes considerably more of the training recipe and supporting resources than most, which matters if you need to audit, reproduce or seriously fine-tune rather than just consume.

10. Phi-4 — Microsoft

A 14B dense model under MIT, and a useful reminder that small and permissive beats huge and restricted for a lot of real work. Punches above its size on reasoning, runs comfortably on hardware you already own, and the licence creates no legal homework.

At a glance

How to actually choose

Ignore the leaderboard for a moment and answer four questions.

Where does it run? If the answer is “my laptop”, the frontier models are irrelevant — a 1.1T-parameter MoE is not running on your machine. Go to Gemma 4 or Phi-4 and be happy.

What is the work? Long agentic coding runs favour GLM-5.2 and the Kimi K2 line. Raw code generation and cost-per-token favour DeepSeek. General-purpose and multimodal favours MiniMax M3.

What licence can you live with? If your legal team needs a name they recognise, that narrows it fast: Apache 2.0 (Qwen, Gemma) or MIT (Phi-4, DeepSeek R1). A community licence may be fine — but decide deliberately, not by accident.

How long will this choice last? On current cadence, assume the model you pick is superseded within a quarter. Build so you can swap it out.

Sources

Rankings and figures above are drawn from these public leaderboards and analyses, captured July 2026:

• BenchLM — best open-source models (BenchAlign v5 index, snapshot 20 July 2026)
• LLM-Stats — open LLM leaderboard
• Vellum — open LLM leaderboard
• Hugging Face — open-source LLMs: benchmarks and licences
• Onyx — best open-source LLMs 2026