Three months ago the open-weight coding tier looked like Kimi K2, GLM-4.6, MiniMax M2, and Qwen3. The thesis then was that the headline parameter counts were nearly decorative and you should pick by active params and post-training. That thesis held up better than the roster: nearly every model on it has since been replaced. Here's the July 2026 bracket, and the same discipline — because the field got faster, not more honest.

The four to know: Kimi K2.7-Code (Moonshot, June 12), GLM-5.2 (Z.ai, June 13), DeepSeek V4-Pro (April 24), and Qwen3-Coder-Next (Alibaba, February 3). All four ship weights you can actually deploy — modified MIT, MIT, MIT, and Apache 2.0 respectively. That's the quietly remarkable part: the permissively-licensed frontier of coding models is now a real, crowded competition, and three of the four come from Chinese labs.

The number that sets your bill: active parameters#

If you self-host, the total parameter count is a vanity metric. What you pay for — GPUs, latency, throughput — tracks active parameters, the slice of the mixture-of-experts that actually fires per token. And the spread here is enormous:

That is more than a 10x gap in serving cost between the cheapest and priciest to run, all wearing the same "open-weight" label. Qwen3-Coder-Next's ~3B active is the outlier that reframes the whole comparison: it's the model you can plausibly run on modest hardware and still get near-frontier SWE-bench numbers out of.

Context: a clean 4x split#

Two of the four target whole-repo work and two don't. GLM-5.2 and DeepSeek V4 offer roughly 1M tokens; Kimi K2.7 and Qwen3-Coder-Next offer 256K (Qwen extrapolates toward 1M, but its native window is 256K). For an agent that needs to hold a large codebase in view, that gap is decisive before you've looked at a single benchmark.

Now the benchmarks — and why you should squint#

Here's the uncomfortable part, and it's the same every quarter: almost every coding score these labs publish is vendor-reported, run on the vendor's own scaffold. Treat the following as claims, not verdicts.

One independent benchmark across four flagship models. The leaderboard screenshot is not evidence; it's marketing with a monospace font.

What to actually do#

Shortlist by constraint, then verify with your own eval — the discipline the coding-agent evaluation playbook exists for.

The roster turned over in a quarter; it will turn over again. What doesn't change is the method: license, context, active-param cost, then your eval. Rank on the vendor's screenshot and you're not choosing a model — you're choosing whose marketing you trust.