Verdict: download Qwen3-Coder-480B-A35B-Instruct if you want one answer. It carries a plain Apache-2.0 licence, and the score most often quoted for it is 69.6% on SWE-bench Verified, the highest reported for an openly licensed coding model. Pick GLM-5.2 instead if your work is long-horizon agent runs across a whole repository, because that is where it is reported to lead. Pick DeepSeek-V3.2 if inference cost is the binding constraint.
That is the whole answer. The rest of this page explains why every "best open-source LLM for coding" list you find says something different, and why the disagreement matters less than it looks.
At a glance
- Default pick: Qwen3-Coder-480B-A35B-Instruct. Apache-2.0, mixture-of-experts, 480B total parameters with 35B active per token.
- Agentic pick: GLM-5.2, MIT licence, open weights published on Hugging Face.
- Cost pick: DeepSeek-V3.2.
- Why the lists conflict: they quote different benchmarks. SWE-bench Verified, SWE-bench Pro and Terminal-Bench are three separate tests, and a model can top one and not the others.
- What does not change week to week: the licence, and whether the weights fit your hardware. Decide on those.
- Last verified: 2026-08-19.
Why do the rankings disagree?
Because they are not measuring the same thing, and almost none of them say so.
SWE-bench Verified is a human-filtered set of real GitHub issue-resolution tasks, described by Modal as 500 tasks. SWE-bench Pro is a harder, separate set. Terminal-Bench scores shell and tool work rather than patch generation. A model tuned for long agent loops can lead on Terminal-Bench and sit mid-table on Verified, and both statements are true at once.
The numbers also move with the harness. One published ranking gives Qwen3-Coder-480B 69.6% on SWE-bench Verified, DeepSeek-V3.2 roughly 70%, MiniMax-M2 69.4% and Kimi K2 71.6% under agentic multi-attempt - and that last qualifier is doing a lot of work. Multi-attempt means the model got more than one try. Comparing a multi-attempt figure to a single-attempt one is not a comparison.
So the honest reading is that the top open-weight coders are clustered within a few points of each other on the same test, and the ordering flips with the harness. Once that is true, the tiebreakers are not benchmark scores.
Qwen3-Coder-480B: the default pick
Qwen's own release notes describe Qwen3-Coder-480B-A35B-Instruct as setting state-of-the-art results among open models on agentic coding, browser use and tool use, comparable to a frontier closed model of its generation.
Two practical reasons it is the default here.
The licence is Apache-2.0. No revenue trigger, no jurisdiction exclusions, no acceptable-use annex to route past your legal team. That matters more than two benchmark points if you ship a product on top of it.
It is everywhere. The model is served by the major inference providers and carries an AWS Bedrock model card as well as an NVIDIA NIM listing, so you can start on an API and move to your own hardware later without changing model.
The catch is size. 480B total parameters is a rented-GPU model, not a laptop model. The 35B active parameter count lowers the compute per token, not the memory you need to hold the weights.
GLM-5.2: pick this for long agent runs
GLM-5.2 is Z.ai's open-weight flagship, published under the MIT licence on Hugging Face. Among the current open trio it is the one reported to lead on SWE-bench Pro and to run at nearly three times the throughput of its peers.
That throughput number is the underrated one. An agent that edits, runs tests, reads the failure and edits again spends most of its wall-clock time waiting on tokens. A model that is slightly weaker per attempt but three times faster can finish more loops in the same hour, and finishing loops is what closes a ticket.
One independent round-up reaches the same split we do, calling GLM-5.2 its overall pick for long-horizon coding agents and repository work, with Qwen3-Coder chosen when Apache-2.0 terms are the priority.
DeepSeek-V3.2: pick this on cost
DeepSeek-V3.2's model card describes a sparse attention mechanism, DeepSeek Sparse Attention, that cuts the cost of long-context work while holding output quality, and claims parity with a frontier closed model of its generation.
For coding specifically, the reason to reach for it is price per solved task rather than peak capability. If your workload is high-volume and repetitive - test generation, migration edits, lint fixes across thousands of files - the cheaper model that is one point behind is the correct choice, and it is not close.
The comparison, on the things that do not move
| Qwen3-Coder-480B | GLM-5.2 | DeepSeek-V3.2 | |
|---|---|---|---|
| Licence | Apache-2.0 | MIT | DeepSeek model licence |
| Shape | MoE, 480B total / 35B active | MoE, open weights on Hugging Face | MoE with sparse attention |
| Strongest reported at | SWE-bench Verified among openly licensed models | SWE-bench Pro, throughput | long-context cost efficiency |
| Runs on a single workstation | No | No | No |
| Best for | shipping a product on open weights | long agent loops over a repo | high-volume, cost-sensitive work |
Can you run any of these on your own machine?
Not at full weights. All three are mixture-of-experts models in the hundreds of billions of parameters, and the memory to hold the weights is the wall, not the compute.
What does run locally is the smaller members of the same families, quantised. That is a genuinely different decision with a different answer, and treating a 480B model and a 27B model as competitors is the most common mistake in this category. If local is your actual constraint, start from what fits in your VRAM and only then compare quality.
What we would not do
We would not pick on a single leaderboard row. The cluster at the top is tight, the harnesses differ, and the ordering has changed more than once this year.
We would also not pick a model whose licence you have not read. "Open-source" is used loosely here: several widely recommended models are open-weight rather than OSI open-source, which means you can download and run them under terms that are not one of the standard permissive licences. Apache-2.0 and MIT are; a bespoke licence with revenue triggers or jurisdiction exclusions is not.
FAQ
Q: What is the best open-source LLM for coding right now?
A: Qwen3-Coder-480B-A35B-Instruct is the best default. It combines an Apache-2.0 licence with the highest commonly reported SWE-bench Verified score among openly licensed coding models, and it is widely available on managed inference.
Q: Is GLM-5.2 better than Qwen3-Coder for coding?
A: For long agent runs across a repository, the reported evidence favours GLM-5.2 on SWE-bench Pro and on throughput. For a single well-scoped patch, and for licence simplicity, Qwen3-Coder is the safer pick.
Q: Why do different sites name different best open-source coding models?
A: Because they quote different benchmarks and different harnesses. SWE-bench Verified, SWE-bench Pro and Terminal-Bench measure different work, and some quoted figures allow multiple attempts while others do not.
Q: Can I run the best open-source coding LLM on a laptop?
A: No. These are mixture-of-experts models with hundreds of billions of total parameters, so the weights alone exceed consumer memory. Smaller quantised models from the same families are the local option.
Q: Is open-weight the same as open-source?
A: No. Open-weight means the weights are downloadable, which is not the same as an OSI-approved licence. Qwen3-Coder is Apache-2.0 and GLM-5.2 is MIT, both permissive; other popular models ship bespoke terms with conditions attached.
Related reading
If your real constraint is running a Qwen model on your own hardware rather than renting it, start with our Qwen 3.8 27B local setup guide. For the GLM family's coding behaviour in more depth, see our GLM 5.3 review on coding and open weights. And for the cheapest way into the DeepSeek family, there is our DeepSeek V4 Flash free setup guide.
Top comments (0)