The best open-weight models in 2026, ranked

A year ago, the interesting question about open-weight models was whether they could do serious work at all. By mid-2026 the question has inverted. The gap between the best open-weight models and the proprietary frontier has narrowed to six points on Artificial Analysis's intelligence index, down from thirteen twelve months earlier. LMArena tells the same story in a different currency: the Elo gap between open and closed models has compressed from roughly 150 points to about 30.

For an institution that wants to run models on its own infrastructure — for data sovereignty, cost control, or regulatory reasons — the shelf is now crowded with credible options. The new problem is choosing between them.

Here is the awkward fact the leaderboards would rather you missed: they disagree about who is winning. Artificial Analysis currently puts GLM-5.2 on top among open-weight models. BenchLM's overall board favours DeepSeek V4 Pro. As recently as May, the same Artificial Analysis index had Kimi K2.6 and Xiaomi's MiMo V2.5 Pro tied for first. Three respected scoreboards, three different champions, in the space of eight weeks.

That churn is the most useful data point in this article. If "best" changes monthly, a selection made on benchmark rank alone will be stale before procurement signs it off. What follows is a ranking, but one ordered by deployability for an organisation that owns its stack — where licence terms, sizing and provenance carry as much weight as raw scores.

A crowded podium

The open-weight frontier in mid-2026 is dominated by a handful of families: DeepSeek's V4 line, Zhipu's GLM-5 series, Alibaba's Qwen 3.5, Moonshot's Kimi K2.6, MiniMax's M3, Meta's Llama 4, and a strong small-model tier led by Google's Gemma and Microsoft's Phi. NVIDIA's Nemotron 3 Ultra sits near the top of the intelligence indices too, and Mistral remains the most prominent European name on the list.

Every one of these can be downloaded and run inside an institution's own perimeter. Under open-weight terms, that is where the similarity ends — the licences, sizes and sweet spots diverge sharply.

The ranking

1. DeepSeek V4 Pro. The most consistent all-rounder across the mid-2026 indices — top or near-top on every board that measures reasoning — released under a clean MIT licence with no revenue or user-count clauses. For an institution that wants frontier-adjacent capability with the simplest possible legal review, this is the current default.
2. GLM-5 family. Artificial Analysis's current #1 open-weight model (GLM-5.2, scoring 51 on the v4.1 index), and the strongest open coding line — GLM-5 posts 77.8% on SWE-bench Verified. Also MIT-licensed. The reasoning variants are the closest thing to a proprietary frontier model that can legally live in your data centre.
3. Qwen 3.5 family. The broadest catalogue: one family, many sizes, from small deployable variants to the 397B flagship, all under Apache 2.0. Strong multilingual performance makes it a frequent shortlist entry for European institutions working across languages. Breadth is the argument — one base-model family can cover several use cases with a shared toolchain.
4. Mistral Large 3. The European flagship: a 675B sparse mixture-of-experts released in December 2025 under Apache 2.0, with a 256K context window, the mid-sized Medium 3.5 (Modified MIT) below it, and a new open-weight family in early access as of July. On raw index scores it trails the Chinese frontier — and it is the only name on this list operating under EU jurisdiction, a provenance that shortens procurement, vendor-risk and regulatory conversations for a European institution in ways a benchmark table can't show.
5. Kimi K2.6. Arguably the strongest agentic performer of the group, and the model that briefly topped the intelligence index in May. The licence is Modified MIT: permissive until an attribution clause activates above 100M monthly users or $20M in monthly revenue. Trivial thresholds for most institutions — but a clause your legal team will want to read, which is exactly the point of ranking on more than scores.
6. MiniMax M3. The newest entrant, released in June 2026: the first open-weight model to combine frontier coding (it leads the open-weight SWE-Bench Pro board at 59.0%), a 1M-token context window and native multimodality. Impressive and young — for a regulated deployment, six months of community scrutiny is worth waiting for.
7. Llama 4 family. Scout's 10M-token context window remains unmatched for extreme long-document work. The catch sits in the licence: a custom agreement with a 700M monthly-active-user cap and restrictions affecting EU deployment. For a European regulated institution, the licence review often ends the conversation before the benchmark review begins.
8. The small tier: Gemma and Phi. Gemma 4 ships under Apache 2.0; Gemma 3 27B runs on a single 16GB GPU, Phi-4 14B on less. These score far below the flagships on intelligence indices — and sit closest to production for a reason. A well-scoped small model on modest hardware is how much of the real deployed work gets done.

What the scores leave out

A ranking compresses; deployment decompresses. Three things the leaderboard position says little about.

Licences bite later. MIT and Apache 2.0 (DeepSeek, GLM, Qwen, Mistral Large 3, Gemma) permit commercial use without conditions. Modified MIT (Kimi, Mistral Medium 3.5) adds thresholds. Custom licences (Llama 4) add geography. The cheapest moment to discover a licence clause is before fine-tuning begins, and the costliest is after production dependency has formed.

Regulation attaches to use, not to weights. From 2 August 2026, the European Commission's enforcement powers over general-purpose AI model obligations take effect, with fines of up to €15M or 3% of worldwide turnover for providers. Open-weight models enjoy partial exemptions — conditional on the model being genuinely open and outside the systemic-risk designation. The Digital Omnibus agreed in June deferred the high-risk system obligations to December 2027 — but deferral is not exemption: the high-risk requirements will land on deployers whichever model sits underneath, and an institution that fine-tunes and redistributes a model can inherit provider-like duties of its own.

Benchmarks measure the general case; your documents are the specific one. A model that scores 51 on an intelligence index and a model that scores 44 can trade places entirely on a portfolio of Polish loan agreements or DACH insurance policies. Frontier proprietary models still hold the edge where tasks are genuinely open-ended and unpredictable — the honest case for the API route remains real. The gap that matters is measured on your own evaluation set, and it is the one number no leaderboard publishes.

Where this lands

The durable finding of mid-2026 is less about any single model and more about the shape of the market: the open-weight frontier now trails the proprietary one by months, not years, and the podium reshuffles monthly. That makes chasing the #1 spot a poor strategy and makes the durable criteria — licence cleanliness, sizing against real hardware, provenance a regulator will accept, performance on your own documents — the actual ranking that matters.

The list above is a starting grid. Which model wins depends on the race your institution is running — and that conversation is a shorter one than most expect.

Related reading:

• Open-weight AI models explained
• Governing open-weight AI models in regulated financial services
• How to choose a base model for your AI project
• Why banks are moving from LLMs to SLMs — and when they shouldn't

Benchmark sources: Artificial Analysis Intelligence Index (v4.1, June–July 2026), LMArena, SWE-bench Verified / SWE-Bench Pro, BenchLM. All figures as reported by those boards at the time of writing.