I keep seeing X is winning open source threads argued from vibes, so I tried the boring version: take the 46 open models we track, group them by the lab that publishes them, and add up only what we already measure per model. No opinion, no benchmark, no forecast — sums and means of readings, with the ordering rule printed above the table.
The rule. A maker’s place is the number of its models in our Heat Score top 10 today (Heat = a published composite of Hub trending, download growth, 30-day downloads and likes, as percentiles among tracked models); ties go to the best-placed model, then to summed 30-day downloads. Every figure is a repository count or an average of per-model scores.
Standings today (29 Sep, 15 makers):
| # | Maker | Top-10 slots | Tracked | Best-placed | Mean Heat, 7-day move | 30-day downloads |
|---|---|---|---|---|---|---|
| 1 | Alibaba Qwen | 3 | 11 | Qwen3 0.6B (#1) | 53, −5 | 60.9 M, +5 % |
| 2 | 3 | 5 | Gemma 4 31B (#2) | 66, +7 | 28.9 M, +9 % | |
| 3 | Zhipu AI | 2 | 5 | GLM-5.3-Flash (#3) | 61, +7 | 8.9 M, +19 % |
| 4 | Meta | 1 | 3 | Llama 3.1 8B (#8) | 51, −14 | 8.8 M, −2 % |
| 5 | OpenAI | 1 | 2 | gpt-oss 120B (#10) | 62, −1 | 11.2 M, −4 % |
| 6 | DeepSeek | 0 | 3 | DeepSeek-V4-Flash (#11) | 48, +1 | 5.5 M, +9 % |
Three things the table says that the threads usually get wrong:
- Breadth and depth are different leads. Qwen holds its slots by volume: eleven tracked models, a mean Heat of 53, and the single hottest model on the site (Qwen3 0.6B). Google holds the same three slots with five models and the highest mean Heat of any maker with more than two models (66). Two ways to lead; the table shows both instead of picking one.
- The movement is not where the slots are. Zhipu’s five models gained seven Heat points on average this week and 19 % in downloads; Meta’s three lost fourteen points, and its best-placed model is a repository from November 2024. Nine of the fifteen makers shipped a tracked model in the last 90 days; Meta, OpenAI, Google, Microsoft and NVIDIA did not.
- Small labs show up as single points. Mistral (two models, +7.5 this week, one Pick of the week), IBM (three Granite models, +8), Liquid AI (three, small but climbing) sit below the slot holders on a rule that rewards top-ten presence — so the page also lists every model per maker with its rank, and says plainly that a maker with one tracked model is judged on that one model.
What it does not say: quality. A download is a download, Heat is attention, and “tracked” means our curated 46 — a lab we do not track is not on the board at all. The full page, refreshed on our five-minute cadence: Maker board — which labs lead open models, by the numbers — AI Compute Radar — and the exact rule with its caveats: Methodology — how every number is made — AI Compute Radar
Happy to add a lab if you think one is missing from the 46; the criterion is an open-weights repo with a measurable GGUF. aspects of your work.