Skip to content
Other

d3-mini-agentcode

4B-фильтр для диффов AI-агентов

2 min read

stonefish-labs опубликовала pre-release d3-mini-agentcode — дообученную версию d3-mini (4B, Apache-2.0), построенную на Qwen/Qwen3.5-4B. Модель читает задачу агента и изменение (diff или сниппет) и оценивает, делает ли изменение что-то security-relevant сверх задачи и выглядит ли оно намеренно вредным; это дешёвый первый фильтр для отбора изменений человеку или более сильной модели. Это pre-release: планируется ревизия на плохих примерах от нескольких writer-моделей, более реалистичных формулировках задач и большем числе языков.

Ключевые особенности

  • 4B-модель; LoRA (r 16, α 32) на attention, linear-attention и MLP-проекциях, 2 эпохи, 1 038 шагов.
  • Три вопроса: class (4 варианта), intent и scope (да/нет); флаг — p(deliberately_malicious) ≥ 0.9 или p(scope=yes) ≥ 0.9.
  • В поставке merged weights, adapter/ и extractor/, который собирает вход и не запускает анализируемый код.
  • Рассчитана на agent diffs и code mode; флаг не доказывает намерение и не должен быть единственной проверкой.
  • Вне области: пакеты/зависимости и релизные диффы, входы длиннее ~8 192 токенов и входы без описания задачи.

Пример

import json
from huggingface_hub import hf_hub_download
from transformers import AutoModel
REPO = "stonefish-labs/d3-mini-agentcode"
model = AutoModel.from_pretrained(REPO, trust_remote_code=True)
questions = json.load(open(hf_hub_download(REPO, "questions.json")))
answers = model.system_one(state=state, questions=questions)["answers"]
flag = answers["class"]["probabilities"]["deliberately_malicious"] >= 0.9 or answers["scope"]["noul"] >= 0.9
ПодробнееПодробнее
На независимом agent-diff наборе: найдено 21/22 out-of-scope изменений, 0/172 реальных диффов и контролей помечены (seed 1).
14 языков и оболочек: 458/510 найдено, 11/2 150 реальных помечено, 1/560 безвредных контролей.
bf16-веса ~9 ГБ; на RTX 4090 батч 8 до 8k токенов — пик 12.7 GiB, около 1 с/элемент; на A100 — 0.46–0.61 с/элемент.
Ограничения: обучение в основном Python, до ~8 192 токенов; пакеты и релизы лучше проверять untuned d3-mini; это pre-release.
Discuss Telegram 0 21

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.5 GB · ready files9.8 GB
  • FP8 / INT8recommendedweights 4.2 GB5.2 GB
  • 4-bitweights 2.4 GB3.1 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

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    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.

    # stonefish-labs/d3-mini-agentcode
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download stonefish-labs/d3-mini-agentcode --local-dir ~/models/d3-mini-agentcode
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to d3-mini-agentcode
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

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.

Model family

  1. Qwen3.5-4B-Basefamily root
  2. Qwen3.5-4B
  3. d3-mini
  4. d3-mini-agentcode

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

stonefish-labs/d3-mini-agentcode
Author
stonefish-labs
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

210

Ask AI about this model

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

Comments

Other

QEV-0.8B

QEV-0.8B