Skip to content
Computer vision

Qwen3.5-0.8B-Python-SoftMoE-v2

a soft MoE Python expert without fine-tuning

2 min read

Qwen3.5-0.8B-Python-SoftMoE-v2 is a compact Python and PEP expert built on top of the 0.87B-parameter dense model Qwen/Qwen3.5-0.8B. Its main feature is weight-space steering (modifying weights directly without gradient training), which isolates a domain direction through an activation scan and injects it into the model weights. Unlike the base version, it includes an embedding-based router that automatically switches Python-related queries to this expert block. It is useful for developers and researchers doing local code analysis and working with specifications.

Key features

  • No fine-tuning: weights are modified via weight-space steering using a gradient-free activation scan.
  • Python specialization: domain prompts are sourced from the official Python EPUB documentation and PEPs (30 chunks).
  • Built-in router: an embedding-margin router automatically activates the expert for on-topic queries.
  • Preserved general skills: HumanEval scores remain on par with the base model without degrading general capabilities.

What you can do

  • Run local experiments and Ivanna Lab pipelines for code generation and PEP checks.
  • Use it as a lightweight component to answer questions about Python syntax and specifications.
  • Compare the behavior of the base Qwen3.5-0.8B and the soft MoE expert on code translation tasks.

Example

When given a query related to Python syntax or PEPs (such as questions about specifications from the official documentation), the built-in embedding router automatically selects this expert over the base model to strictly follow style guidelines.

Read moreDetails
• Base model: Qwen/Qwen3.5-0.8B (0.87B parameters)
• Mode and parameters: steer, α = 0.1, layers [15, 16, 17, 18, 20, 21, 22, 23]
• Benchmark: code2text_python (CodeXGLUE) BLEU-4 = 1.0534
• Method: Weight-Space Programming (MIT CSAIL, 2026)
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.6 GB · ready files2.3 GB
  • FP8 / INT8recommendedweights 0.8 GB1.4 GB
  • GGUF Q8_0weights 0.8 GB · ready files1.4 GB
  • 4-bitweights 0.5 GB1.0 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 2060

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.

    # NikitaLazarev/Qwen3.5-0.8B-Python-SoftMoE-v2
    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 NikitaLazarev/Qwen3.5-0.8B-Python-SoftMoE-v2 --local-dir ~/models/Qwen3.5-0.8B-Python-SoftMoE-v2
    # The model is in ~/models. How to run it is in the model card (link above).
  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. Qwen3.5-0.8B-Basefamily root
  2. Qwen3.5-0.8B
  3. Qwen3.5-0.8B-Python-SoftMoE-v2

New posts about versions of Qwen3.5-0.8B-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 9All on Hugging Face
Merges 1All on Hugging Face

About the model

NikitaLazarev/Qwen3.5-0.8B-Python-SoftMoE-v2
Author
NikitaLazarev
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments