Open-source model roundup: July 2026
Open weights are reaching frontier scale — which is exactly why 'open' no longer means 'runnable on your laptop'.
Updated 31 Jul 2026 fast-moving — check a current source
Release names and dates gathered from model trackers in late July 2026; parameter counts are as published by the labs. Benchmark claims are the labs' own — verify on your own tasks.
The open-weight scene had a loud July, and the releases point somewhere genuinely new: the frontier and the open-weight world are converging on size. That is good news for capability and awkward news for anyone who assumed “open” meant “I can run this”.
Notable this month
- Kimi K3 (Moonshot AI, 16 July) — a 2.8-trillion-parameter open-weight mixture-of-experts model with native vision and a one-million-token context window.
- Qwen3.7 Flash (Alibaba, 27 July) — the latest in the Qwen line, aimed at speed.
- Still current from recent months: GLM-5.2 from Z.ai, DeepSeek V4 for reasoning, and Gemma 4 at the small end for laptops and edge devices.
The size problem
A 2.8-trillion-parameter model is open in the sense that matters legally — you can download the weights, inspect them, fine-tune them, and run them without asking permission. It is not open in the sense most people mean when they say they want to run a model locally. Serving a model at that scale needs a cluster, not a workstation.
This is worth naming plainly because the word “open-source” is doing two different jobs:
- Open as sovereignty — no vendor dependency, no data leaving infrastructure you control, no unannounced model changes. Achievable at frontier scale, if you have the hardware budget.
- Open as accessibility — a good model on the machine you already own. That still means the small end of the range, and the small end is where Gemma 4 and the compact Qwen builds matter.
Both are legitimate. They just have almost nothing to do with each other in practice, and a release that serves one may be useless for the other.
If a roundup makes you want to download something, check the parameter count against your actual hardware before you check the benchmark scores. The best model you cannot run is worth less than the mediocre one you can.
How to evaluate a release
Unchanged, and still the only advice that survives contact with a new leaderboard: pull a candidate, run it on your tasks, and weigh quality against the hardware it needs — including the energy trade-offs of serving it yourself. For last month’s picture, see June’s roundup.
Sources
Everything above was checked against these on 31 Jul 2026. Providers change things without notice — if a detail matters to a decision, follow the link.