z3-mini soup preview
A preview of the Zils soup method: a plain, uniform weight average of three open decision models built on Qwen3.5-4B. This is not a z3-mini release. z3-mini v1 remains the current release. A soup only becomes the next z3-mini after it passes the release gate.
What it is
The Zils network builds a "soup" every day: it averages the bodies of different Qwen3.5-4B decision fine-tunes, adding a model only when the average scores better with it. This checkpoint is the soup from the first daily round. It contains:
| Ingredient | Author | Licence |
|---|---|---|
torchcast-decision-4b @ 492bfb1 |
torchcast-ai | Apache-2.0 |
pplx-decider-flash-v1.2 @ 9ff7cf9 |
Perplexity | Apache-2.0 |
d3-mini @ df99bf4 (z3-mini v1) |
vllm-sr | Apache-2.0 |
- Body: the uniform average (computed in fp32, saved in bf16) of the three
models'
language_model.*andvisual.*weights. - Answer layer, prompt and runtime: unchanged from d3-mini. The answer layer is byte-identical to d3-mini's.
- Soup hash:
063e05dc12c15db265de911499582549e6acf02840dccc38dc1edb1482689cd6
How it scored
The scores below come from Zils' own private exam: 3,000 decision questions with labels from real data, across five categories. They are not public benchmark results, and other evaluations may rank these models differently.
| Model | Score (0–100) |
|---|---|
| This soup | 74.3 |
| torchcast-decision-4b alone | 72.7 |
| pplx-decider-flash-v1.2 alone | 71.4 |
| d3-mini (z3-mini v1) alone | 69.7 |
| Category | Soup |
|---|---|
| Arts | 76.9 |
| Language | 80.1 |
| Reasoning | 50.7 |
| Retrieval | 87.3 |
| Tools | 84.1 |
On this exam the soup beats each of its ingredients alone: torchcast by 1.6 points, pplx by 2.9 and d3-mini (z3-mini v1) by 4.6.
Use
It runs exactly like d3-mini v3.1.0: the same bundled runtime files
(d3_runtime.py, d3_server.py, …), prompt format and /v1/systemone API. Follow the
d3-mini v3.1.0 usage instructions
and point them at this repository.
Limits
- Text decisions only. The exam is text-only and doesn't measure image, video or bounding-box quality. Image and box behaviour may be worse than any one ingredient's.
- Possible benchmark overlap. The ingredients are third-party models, and their training data may overlap with public benchmarks.
- Not a release. It hasn't been through the Zils release gate (sealed release draws, the confirmation draw and the adapter suite). Adapters trained on z3-mini v1 are not guaranteed to work on it.
Licence and credits
Released under Apache-2.0. The licences and NOTICE files of all three ingredients
are included unchanged: LICENSE, NOTICE (d3-mini), LICENSE-torchcast,
NOTICE-torchcast, LICENSE-pplx, NOTICE-pplx, and LICENSE-qwen-Apache-2.0.
The Qwen3.5-4B base is by the Qwen team / Alibaba Cloud (Apache-2.0). Thanks to
torchcast-ai, Perplexity and vllm-sr for publishing their models openly.
Modifications: the language_model.* and visual.* weights are the uniform
average of the three ingredients' weights. Every other file (answer layer,
tokenizer, prompt and runtime) is d3-mini's, unchanged. No affiliation with or
endorsement by the ingredients' authors, Qwen or Alibaba Cloud is implied.
- Downloads last month
- 75