If you want the best open-weight coding model you can self-host in September 2026, the honest answer is that it depends on your GPU, not the leaderboard. For raw capability, DeepSeek V4 Pro tops the open-weight SWE-bench Verified board; for a model you can actually download, own, and ship commercially, Qwen3-Coder under Apache-2.0 is the pick; and if you have exactly one 24GB GPU, Qwen3-Coder 30B-A3B is the best coder that fits.

Here is the ranking, open-weight only, with what each one takes to run:

  1. DeepSeek V4 Pro — highest raw coding scores of any open model, but a 1.6T-parameter giant that is API-or-datacenter-only in practice.
  2. Qwen3-Coder 480B-A35B — the best coder you can fully own, under the most permissive license here (Apache-2.0).
  3. Kimi K3 (Moonshot) — the strongest open model for agentic and long-horizon coding.
  4. GLM-5.2 (Zhipu) — frontier coding at a fraction of the cost, MIT-licensed, popular inside coding agents.
  5. Devstral 2 (Mistral) — a model purpose-built to be a coding agent, more tractable to self-host than the 1T-class giants.
  6. Qwen3-Coder 30B-A3B — the best model that fits on a single consumer GPU.
  7. gpt-oss-120b / 20b (OpenAI) — Apache-2.0 generalists with a genuinely good "runs on one card" story.

This list is open-weight only on purpose. If you want the combined picture including proprietary APIs, we keep a separate coding-model leaderboard. And a model is not an agent — the tool that drives it (the editor, the loop, the sandbox) matters as much as the weights, which is why we rank those separately in the AI coding agent ranking. This piece is about the weights you can download.

The frontier you can download (but probably can't fit)#

The top of the open-weight table is dominated by trillion-parameter Mixture-of-Experts models. They are open in the license sense and closed in the practical sense: you can have the weights, you just can't fit them on anything you own. (One caveat that just changed: the commons where you download most of these — Hugging Face — is now being acquired by Nvidia, so pin the exact versions you ship and mirror the weights you depend on.)

DeepSeek V4 Pro is the current open-weight leader on the SWE-bench Verified tracker, landing around the low-80s and effectively tying the best-scoring proprietary models on that board. It is a 1.6T-parameter MoE (roughly 49B active per token) shipped under a clean MIT license with a 1M-token context. The catch is size: the FP8 checkpoint is on the order of 893GB. That is not a home-lab number — it is a multi-node serving job. For almost everyone, DeepSeek V4 is something you consume through an API or a serving provider, not something you host.

Qwen3-Coder 480B-A35B is the one I'd actually reach for if the goal is to own the model. Per Qwen's own release, it is a 480B MoE with 35B active parameters, a 256K native context (extendable toward 1M), and a headline 66.5% Pass@1 on SWE-bench Verified — state-of-the-art among open models at launch and still a top-tier agentic coder. Crucially, it is Apache-2.0: no field-of-use restrictions, no scale conditions, genuinely commercial-friendly. It still wants a multi-GPU box, but 480B/A35B is far more approachable to serve than a 1.6T monster, and it's the model most self-hosting teams standardize on.

Kimi K3 from Moonshot is the community favorite for agentic and long-horizon work — released open-weight on July 16, 2026, it's the largest open-weight model publicly available at 2.8T parameters (a sparse MoE) with a 1M-token context, and it leads sustained-coding boards like SWE Marathon. It ships under a Modified MIT license that adds conditions for very large-scale commercial deployments, so read that clause before you build a product on it. At 2.8T parameters it is firmly a cluster-or-API model — you're pulling these weights to serve them on rented infrastructure, not on your desk.

GLM-5.2 from Zhipu is the value play at the frontier: a ~753B MoE with a 1M context under MIT, and it's the model a lot of people quietly run behind Claude-Code-style agents because it's cheap and strong at coding. Zhipu has been shipping fast — GLM-5.3 landed in mid-August 2026 — so check which checkpoint is current when you pull weights, and note that the flagship and the smaller "Flash" variant have sometimes shipped under different licenses.

The models you can actually self-host#

This is where the piece earns its title. Below the trillion-parameter tier sit models that fit real hardware.

Qwen3-Coder 30B-A3B is the standout. It's a 30B MoE that activates only 3B parameters per token, and at 4-bit it needs roughly 17-22GB of VRAM (Unsloth's local guide puts Q4_K_M around 21.9GB) — comfortably inside a single 24GB card like a used RTX 3090 or a 4090. You get the Qwen coding lineage and Apache-2.0 on a machine you might already own. For most solo builders, this is the local coding model.

gpt-oss from OpenAI is the other easy win. Per OpenAI's release, gpt-oss-20b (21B total, ~3.6B active) fits in about 16GB of VRAM, and gpt-oss-120b (117B total, ~5.1B active) fits on a single 80GB card like an H100 or MI300X. Both are Apache-2.0, ship MXFP4-quantized with a 128K context, and expose low/medium/high reasoning effort. The 120b-on-one-card story is the most convincing "serious model, one GPU" option on this list.

Devstral 2 from Mistral is built specifically to run inside coding agents. VentureBeat reported the Devstral 2 launch at ~72% on SWE-bench Verified for the 123B flagship, with Devstral Small 2 (24B) in the high-60s and runnable on a single RTX 4090 or a 32GB Mac. It's released under a modified-MIT license that permits commercial self-hosting. If you want a mid-size model tuned for tool-calling and repo-scale edits rather than chat, this is the one.

A note on Codestral: Mistral's original 22B code model is great at fill-in-the-middle autocomplete across 80+ languages, but its license is non-production for the base weights. For a truly open coding model from Mistral, Devstral is the answer.

How to actually run it: VRAM tiers#

Pick your tier by the hardware you have, then pick the best model in it.

The rule of thumb: for anything in the trillion-parameter tier, a hosted API almost always beats self-hosting on cost unless you're running rented GPUs at high, sustained utilization.

Licenses matter more than the last benchmark point#

If this is for a business, the license column outranks a two-point SWE-bench gap. The genuinely unrestricted options here are Apache-2.0 (Qwen3-Coder, gpt-oss) and MIT (DeepSeek, GLM). Kimi K3 and Devstral use modified-MIT variants that are permissive for most users but attach conditions for very large-scale deployments — fine for a solo product, worth a lawyer's glance at scale. Codestral's base weights are the outlier to avoid for production.

How these were ranked#

Open-weight universe only — models whose weights you can download, no proprietary APIs. The order weights two published coding benchmarks — SWE-bench Verified (agentic, real GitHub bug-fixes) and LiveCodeBench (contamination-free competitive coding) — against how practical the model is to actually self-host, which is a function of license permissiveness and VRAM footprint. That's why Qwen3-Coder can outrank models with a slightly higher raw score: a model you can legally own and realistically serve beats one you can only rent.

Benchmark numbers move weekly and vendors quote favorable configs, so treat every figure as directional and click through to the primary leaderboards linked above before you commit. If you'd rather skip hosting entirely, our LLM API pricing comparison covers what these same models cost per token from providers.

Bottom line: start from your GPU. One 24GB card, run Qwen3-Coder 30B-A3B. One 80GB card, run gpt-oss-120b. A rack or a rented node, run Qwen3-Coder 480B or GLM-5.2. Only the truly giant frontier models (DeepSeek V4, Kimi K3) belong on someone else's servers.