The Zen family
45 open-weight models across language, code, vision, image, audio, and retrieval — built by Zoo Labs Foundation. Free to self-host, or managed on Hanzo Cloud. Benchmarks here are UPSTREAM-reported for the open ecosystem Zen builds on; only Enso is Hanzo-measured end-to-end.
Generations, and what each one is for
A generation is a training run, not a marketing tier: newer does not mean the older ones stop working, and a small model from an older generation is often the right call. The large ones are Zen MoDE — Mixture of Diverse Experts. Open any generation for specs and prices.
Where open weights stand
44 open-weight models across the field, by benchmark, so you can see what running your own hardware costs you in capability before you commit to it. A figure is what its vendor published unless it is tagged Hanzo, which means we ran it ourselves. Toggle the provenance to see which is which.
| Model | GPQA-Diamond | Source | $/MTok out |
|---|---|---|---|
| kimi-k2.6 | 89.1 | Vals AI | $2.71 |
| qwen3.5-397b-a17b | 88.4 | LLM Stats | $2.04 |
| nemotron-3-ultra-550b-a55b | 86.1 | Vals AI | $1.54 |
| glm-5.2 | 85.6 | Vals AI | $3.73 |
| glm-5.1 | 84.5 | Vals AI | $3.63 |
| kimi-k2.5 | 84.1 | Vals AI | $1.69 |
| glm-5 | 83.3 | Vals AI | $2.07 |
| nemotron-3-super | 82.7 | LLM Stats | $0.4 |
| minimax-m2.5 | 82.1 | Vals AI | $0.76 |
| mimo-v2.5 | 81.6 | Vals AI | $0.24 |
| deepseek-v3.2 | 80.3 | Vals AI | — |
| deepseek-v3.2-exp | 79.9 | DeepSeek-V3.2-Exp model … | — |
| deepseek-4-flash | 76.9 | Hanzo | $0.2 |
| deepseek-v4-pro | 75.3 | Hanzo | $2.5 |
| deepseek-r1 | 71.5 | DeepSeek-R1 model card (… | — |
| llama-4-maverick | 69.4 | Vals AI | $0.75 |
| llama-3.3-70b-instruct | 50.5 | LLM Stats | — |
Vendors report on their own harness; Hanzo measures everyone on one. Where both exist the gap is the harness talking — not the model getting better. Hover a source for its exact provenance.
Why the two labels. A score is only comparable to another score run the same way, and most published figures were not. So the table keeps the distinction rather than averaging it away: a figure we ran on our own harness says Hanzo, and a figure a vendor published says so too. Enso is the family we measure end to end on one common harness — see it on the Enso page. Everything else on this page is cited, not claimed.