What changed this month
- MiniMax M3 now leads open-weight on SWE-Bench Pro at 59.0% (launched 1 June 2026), just ahead of Kimi K2.6's 58.6% from April. Remember: SWE-Bench Pro is the harder set — scores live in the mid-to-high 50s.
- DeepSeek V4-Pro still leads open-weight on SWE-Bench Verified at 80.6% — a different, easier benchmark where scores sit in the low 80s. MiniMax M3 is right behind at 80.5%.
- M3 is the first open-weight model to pair frontier coding with a 1M-token context window and native multimodality — image and video input — in one set of downloadable weights.
- The front rank is now five-deep: MiniMax M3, GLM-5.1, Kimi K2.6, DeepSeek V4-Pro and Qwen3-Coder-Next. The practical decision has shifted from "open vs closed" to "which open-weight model fits my stack".
There are two different benchmarks in play and they are not comparable. SWE-Bench Pro scores in the ~50s; SWE-Bench Verified scores in the ~80s. A model can lead one and trail the other — DeepSeek V4-Pro and MiniMax M3 swap places depending on which set you read. Always check which benchmark a "leader" claim refers to before you commit hardware to it.
The two benchmarks, kept separate
Most of the confusion around open-weight coding leaderboards comes from mixing up the two SWE-Bench variants. We are going to keep them apart for the whole article, and you should too when you read anyone else's comparison.
SWE-Bench Pro is the harder evaluation — real, gnarly software-engineering tasks where even frontier systems struggle to clear 60%. Here the open-weight order in June 2026 is MiniMax M3 at 59.0%, Kimi K2.6 at 58.6%, and GLM-5.1 at 58.4%. For context, when GLM-5.1 released in April it scored above GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%) on this set — a genuine moment for open weights, because it meant a downloadable model was beating named closed frontier APIs on the hard benchmark.
SWE-Bench Verified is the easier, cleaner set — scores cluster in the low 80s. Here DeepSeek V4-Pro leads the open-weight pack at 80.6%, with MiniMax M3 at 80.5%, Kimi K2.6 at 80.2% and the older GLM-5 at 77.8%. The overall leaderboard, including closed models, is still topped by GPT-5.5 at 88.7% — so on the easier set, closed frontier remains ahead, even as the open-weight field closes in. We dug into that gap-closing trend in our piece on how open weights closed the gap on GPT-5.5.
Benchmark your own repository before you trust any leaderboard. SWE-Bench tasks skew towards Python and well-known open-source projects. If your codebase is a TypeScript monorepo in Bengaluru or a legacy Java estate in Manchester, run a small held-out set of your own tickets through two or three candidates before you pick. The leaderboard ranking and your ranking will not always agree.
The front rank, side by side
Here is the open-weight coding field as it stands in June 2026. Scores are labelled by benchmark; "best for" is our read on where each model earns its keep in a self-hosted stack.
| Model | SWE-Bench Pro | SWE-Bench Verified | Context window | Modality | Best for |
|---|---|---|---|---|---|
| MiniMax M3 | 59.0% (open-weight lead) | 80.5% | 1M tokens | Text + image + video | Long-context, multimodal coding agents in one model |
| Kimi K2.6 | 58.6% | 80.2% | Long-context | Text | Strong all-rounder on the hard set; mature tooling |
| GLM-5.1 | 58.4% | — | Long-context | Text | Beat GPT-5.4 and Opus 4.6 on Pro at release; hard-task reasoning |
| DeepSeek V4-Pro | — | 80.6% (open-weight lead) | Long-context | Text | Top open-weight on Verified, LiveCodeBench and Codeforces |
| Qwen3-Coder-Next | — | — | Long-context | Text | Front-rank contender; broad ecosystem and fine-tune support |
A dash means we are not citing a verified figure on that benchmark for that model — not a zero. Older GLM-5 scored 77.8% on SWE-Bench Verified for reference.
MiniMax M3: the multimodal long-context bet
The reason M3 is the headline this month is not only that 59.0% on SWE-Bench Pro. It is the combination. M3 is the first open-weight model to put frontier coding, a 1M-token context window and native multimodality — image and video input — into a single set of weights you can download and run yourself. The weights were set to be available for download within roughly ten days of the 1 June launch.
That combination matters for a specific class of work. A coding agent that can read a 1M-token codebase context in one pass, take a screenshot of a broken UI or a Figma frame as input, and reason over a short screen-recording of a reproduction is a meaningfully different tool from a text-only model wired to a separate vision service. For teams building front-end agents, design-to-code pipelines or QA-from-recording workflows, M3 collapses two or three models into one.
VentureBeat, citing MiniMax, reported that M3 "eclipses GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5–10% of the cost." We would treat that as a vendor-and-press claim rather than an independently established fact — the framing is MiniMax's, and "key benchmark performance" does not specify which benchmark. The verified, separable figures are the ones in our table above. Still, even discounted, the cost story is the point: open weights you self-host change the unit economics, which is exactly why this matters for budget-conscious teams. Our deeper explainer on running it lives at MiniMax M3 as open-weight frontier.
If your roadmap includes any vision-in-the-loop coding — design-to-code, visual regression triage, or agents that read screen recordings — shortlist MiniMax M3 first. It is the only front-rank open-weight model that handles image and video natively, so you avoid stitching a separate vision model into the agent loop.
DeepSeek V4-Pro: the breadth play
If MiniMax M3 wins on the hard SWE-Bench Pro set, DeepSeek V4-Pro answers on breadth. It leads open-weight models on SWE-Bench Verified at 80.6%, and it also tops LiveCodeBench (93.5) and Codeforces (rating 3206) among the models evaluated — including closed frontier APIs. For a team whose work is more "competitive-programming-shaped" — tight algorithmic problems, contest-style tasks, raw code generation rather than long multi-file engineering — V4-Pro is the model to beat.
It is also the model with the clearest self-host cost story so far. We worked through the GPU maths in DeepSeek V4-Pro self-host break-even on 8×H100, and the headline is that at steady, high throughput the fixed hardware cost is amortised quickly enough to beat per-token API pricing. The catch is "steady and high" — which brings us to the decision that actually matters.
So which one do you self-host?
The benchmark leaders are interesting, but the real decision for a builder in Pune or Leeds is operational, not academic. Three questions decide it.
- What shape is your work? Hard, multi-file engineering tickets point to MiniMax M3, Kimi K2.6 or GLM-5.1 (the SWE-Bench Pro leaders). Algorithmic and contest-style code generation points to DeepSeek V4-Pro. Vision-in-the-loop points to MiniMax M3 alone.
- What is your throughput? Self-hosting only pays when GPUs are kept busy. Spiky, low-volume traffic almost always favours a hosted API. We laid out the unit economics in AI inference cost economics in 2026 — read it before you sign a hardware lease.
- What are your constraints? Data residency, no API lock-in and predictable monthly cost are the three reasons builders self-host even when the napkin maths is close. For regulated work in either market, keeping the weights and the data on infrastructure you control is often the deciding factor regardless of price.
We are not going to re-explain self-hosting mechanics here — the break-even and inference-cost pieces above cover the GPU sizing, batching and amortisation in depth. The point of this article is the model choice, and on that the picture is clearer than it has been all year. May's field already looked crowded; our May open-weight coding shoot-out ranked four serious contenders. M3 has simply raised the ceiling again, a month later.
"Open-weight" is not the same as "open-source". Several of these models ship under licences with usage restrictions — commercial-scale, competitor or jurisdiction clauses. Before you build a product on any of them, read the actual licence, not the launch tweet. The right model under the wrong licence is the wrong model.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The bottom line
For the first time, the most capable coding model you can run on your own hardware is genuinely competitive with the best closed APIs on the hard benchmark — MiniMax M3's 59.0% on SWE-Bench Pro sits above the GPT-5.4 and Opus 4.6 figures GLM-5.1 was already beating in April. On the easier SWE-Bench Verified set, closed frontier still leads (GPT-5.5 at 88.7%), but the open-weight field is within touching distance and closing.
If you are choosing today: pick MiniMax M3 for long-context and multimodal coding, DeepSeek V4-Pro for breadth and algorithmic work, Kimi K2.6 and GLM-5.1 as strong hard-task all-rounders, and keep Qwen3-Coder-Next on the shortlist for its ecosystem and fine-tuning support. Then benchmark your own repository before you commit a single GPU. The leaderboard is a starting point, not the answer.
More open-weight coverage is collected under our open-source news section.