Skip to content
Language models

z3-mini soup preview 2

a merge of three Qwen3.5-4B decision-making models

2 min readupdated

A daily preview checkpoint from the Zils network created via a model soup (uniformly averaging the weights of several trained models into one). Built on the 4.54B-parameter Qwen3.5-4B architecture, it merges the language block weights from three models trained by Zils miners for decision-selection tasks. Compared to the base d3-mini (z3-mini v1) release, this build delivers substantially higher probabilistic accuracy on private test sets. It fits developers looking for a compact model for text-based classification and decision-making via a d3-mini-compatible API.

Key features

  • Uniformly averages the language_model.* weights of three fine-tuned models (zils-mini-hm4, zils-mini-decision-t2avg, and zils-mini-m1113b) in fp32 and saves the result in bf16.
  • Keeps the vision tower, output head, tokenizer, and prompt format unchanged from d3-mini.
  • Bakes a calibrated temperature of 1.36 into model.json, tuned by the network on a held-out validation set.
  • Fully compatible with the d3-mini v3.1.0 launch format and /v1/systemone API.

What you can do

  • Use it as a drop-in replacement for d3-mini in existing pipelines built around d3_runtime.py and the /v1/systemone endpoint.
  • Deploy it in automated text scenario scoring and decision-making pipelines where probability calibration (Brier skill) matters.
  • Benchmark its performance against single models (torchcast-decision-4b, Matilda-Decision-v1-Mini) on your own text datasets.

Example

Run the model through the standard d3-mini v3.1.0 runtime (d3_runtime.py and d3_engine.py): point the launch config to this repository and send text decision-making requests in the d3-mini prompt format to the /v1/systemone endpoint.

Read moreDetails
Zils private test set (25,008 questions, Brier skill 0–100): 58.1 vs. 46.9 for d3-mini (z3-mini v1).
November 16, 2026 round: 56.8 vs. 48.3 for torchcast-decision-4b, 47.6 for Matilda-Decision-v1-Mini, and 43.9 for pplx-decider-flash-v1.2.
Limitations: this is a preview build that has not gone through final Zils release gating; adapters from z3-mini v1 may not work; only text-based decisions were evaluated; training data from the ingredient models may overlap with public benchmarks.
License: Apache-2.0.
Discuss Telegram 0 24

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.5 GB · ready files9.9 GB
  • FP8 / INT8recommendedweights 4.2 GB5.3 GB
  • 4-bitweights 2.4 GB3.2 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 3080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # zils-ai/z3-mini-soup-preview-2 — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve zils-ai/z3-mini-soup-preview-2 --max-model-len 8192 --quantization fp8
    Add more LoRAs to z3-mini-soup-preview-2
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

  1. Qwen3.5-4B-Basefamily root
  2. Qwen3.5-4B
  3. z3-mini-soup-preview-2

New posts about versions of Qwen3.5-4B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 12All on Hugging Face
Merges 1All on Hugging Face

About the model

zils-ai/z3-mini-soup-preview-2
Author
zils-ai
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

240

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments