The Vibe Father
Roundup
Two years ago, "open-weight coding model" meant "the one you apologize for." In mid-2026 it means a model posting a 78.7 on SWE-bench Verified, another one selling tokens at 87 cents per million output, and a serious argument that for most builder-seat work the open models are simply the rational choice. This roundup covers the open-weight models on our board, what open-weight actually means, and — honestly — where the gap to the closed flagships still lives.
First, what "open-weight" means
Open-weight means the model's weights are published: you can download them, run them on your own hardware, fine-tune them, and serve them yourself or through any inference provider you trust. It does not automatically mean "open source" in the classic sense, and it does not mean "do anything you want" — each model ships under its own license, and the commercial terms vary from genuinely permissive to conditional. Before you build a business on one, read its actual license; we deliberately do not summarize specific license terms here because they change and the details matter.
Why it matters is threefold. Price pressure: open weights mean multiple providers compete to serve the same model, which is a big part of why the cheapest rows on our board are all open models. Privacy and control: you can run them inside your own perimeter, with no code leaving the building. And no rug-pulls: a closed model can be deprecated, repriced, or quietly changed under you; weights on your own disk cannot.
The open-weight board
All scores from public sources, live at /benchmarks. Where a score is not yet published, we say so — we never fill in a number.
| Model | SWE-bench | Terminal-Bench | LiveCodeBench | API price (in/out per M) | Context |
|---|---|---|---|---|---|
 GLM 5.2 | 78.7 | not published | not published | — | — |
 DeepSeek V4 Pro | 77.6 | not published | 87.5 | $0.435 / $0.87 | 1M |
 Qwen3.7 Max | 77.3 | not published | 87.1 | $2.50 / $7.50 | 1M |
 Kimi K2.6 | 76.7 | 66.7 | 86.8 | — | — |
| Kimi K2.7 Code | not published | not published | 82.1 | $0.95 / $4 | 262k |
 MiniMax M3 | not published | not published | 82.2 | $0.30 / $1.20 | 1M |
GLM 5.2 — the new SWE leader
Z.ai's GLM 5.2 currently holds the best open-weight SWE-bench Verified score at 78.7 — a number that would have led our entire board eighteen months ago. Its Terminal-Bench and LiveCodeBench scores are not yet published, so on our index it is renormalized around the one score that exists rather than punished for the gaps. If the missing scores land in line with the SWE number, this is the open model to beat.
DeepSeek V4 Pro — the complete package
DeepSeek V4 Pro is the open model we recommend most, because it pairs near-top scores with the most aggressive pricing anywhere: 77.6 SWE-bench Verified, 87.5 LiveCodeBench, 1M context, at $0.435 in and $0.87 out via API. That is a published, verifiable, flagship-adjacent scorecard for roughly the price of a rounding error. Terminal-Bench is the missing datapoint. Full analysis in our DeepSeek V4 Pro review.
Qwen3.7 Max — the fast one
Qwen3.7 Max posts 77.3 SWE and 87.1 LiveCodeBench while streaming at 204 tokens per second — the fastest model on our entire board, open or closed. For agent loops where iteration cadence compounds, that throughput is a feature no closed flagship matches.
Kimi and MiniMax — the value tier
Moonshot's Kimi K2.6 is the only open model here with a complete published slate — 76.7 SWE, 66.7 Terminal-Bench, 86.8 LiveCodeBench — which makes it the most honest datapoint in the tier, and that 66.7 Terminal-Bench candidly shows where open models still trail on agentic shell work. Its successor K2.7 Code ($0.95/$4) has only its 82.1 LiveCodeBench published so far; see our Kimi K2 review for the family history. MiniMax M3 is the price floor of the entire market at $0.30/$1.20 with an 82.2 LiveCodeBench and everything else unpublished.
The honest gap
Now the part advocates skip. On SWE-bench Verified — the benchmark closest to real repository work — the best open score is GLM 5.2's 78.7. Claude Fable 5 sits at 95.0 and Opus 4.8 at 88.6. That is not a rounding error; it is a ten-to-sixteen-point canyon, and on hard multi-file tasks you feel it. The open models have effectively caught the closed mid-tier (Sonnet 5's 85.2 is the nearest closed builder) and interestingly cluster right at the closed models' LiveCodeBench level. But nobody's weights-on-disk currently plan like a Fable.
Self-hosting: the fine print
One expectation worth calibrating before you provision GPUs: "you can run it yourself" and "you should" are different claims. Frontier-scale open models are enormous, and serving one at production quality means serious hardware, an inference stack to operate, and throughput that may well trail the model's own first-party API — which is often startlingly cheap precisely because the weights are open and providers compete. For most teams, the rational default is using open models through hosted APIs and treating self-hosting as the option you exercise when a requirement forces it: data that cannot leave your perimeter, a compliance regime, a fine-tune you cannot run anywhere else, or simple insurance against provider changes.
The insurance framing is underrated. Even if you never self-host, the fact that you could disciplines the market: an open model's API price cannot drift far from the cost of serving it yourself, and the model cannot be deprecated out from under a workflow you depend on. Some of the value of open weights accrues to you without ever downloading them.
Bottom line
Our playbook: open models in the volume seats — builders, scouts, drafting — where DeepSeek and Qwen deliver 90% of a closed builder at 10–30% of the price, and a closed flagship where depth decides outcomes. If your constraint is budget, start with our cheapest models roundup; if it is privacy, the fact that this tier exists at this quality means self-hosting no longer requires a quality apology. Live scores at /benchmarks.