Skip to content
Other

qwen3.5-4b-prune_width_ffn15_lora_kd

compressed LLM with knowledge distillation

2 min read

This model is the result of a language compression and knowledge distillation experiment (where a smaller model is trained to mimic the responses of a larger teacher model). It is based on a pruned version of mrysh/qwen3.5-4b-prune_width_ffn15 with its MLP (multilayer perceptron) width reduced by 15%, which was then fine-tuned using LoRA (a small adjustable add-on over the base model) on a 20-million-token corpus. After training, the LoRA weights were merged back into the model, resulting in a standard checkpoint with a smaller disk footprint and the architecture of the original pruned model. The model is useful for researchers and developers looking for faster and more compact LLM versions to run on limited resources.

Key features

  • Knowledge distillation: KL-divergence was used to align the token distribution of the pruned model with the original Qwen/Qwen3.5-4B.
  • High-quality training data: Utilized texts from FineWeb-Edu, OpenAssistant dialogues, and specialized prompts from the BFCL and MedQA benchmarks.
  • Reduced size: Weighs 7.74 GB on disk (compared to 8.41 GB for the original) while maintaining the identical structure of the pruned version.
  • High speed: Single-user generation speed reaches 79.6 tokens per second on an L40 GPU via vLLM.

What you can do

  • Pipeline integration: Use the model as a lighter replacement for the base Qwen3.5-4B for text tasks.
  • Feature testing: Test function-calling performance using the BFCL benchmark.
  • Service deployment: Deploy the model via vLLM for fast inference in bots and applications.

Example

Load the model the standard way, just like the base Qwen3.5-4B, taking into account the modified layer widths:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mrysh/qwen3.5-4b-prune_width_ffn15_lora_kd"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
Read moreDetails
• Benchmarks: BFCL (total) — 86.4% (original is 88.0%); MedQA — 61.5% (original is 70.8%, pruned without distillation is 56.5%); Hiring bias — 81.9%.
• Hardware requirements: Disk size 7.74 GB, speed 79.6 tokens/s (vLLM, L40).
• Training details: LoRA (rank 64, alpha 128) on all linear layers of the text decoder, 606 optimizer steps, LR 0.0001 with cosine decay.
• Limitations: Tests were conducted with thinking mode disabled and greedy decoding; results reflect behavior on specific training datasets.
Discuss Telegram 0 21

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 7.2 GB · ready files8.4 GB
  • FP8 / INT8recommendedweights 3.6 GB4.5 GB
  • 4-bitweights 2.0 GB2.7 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.

    # mrysh/qwen3.5-4b-prune_width_ffn15_lora_kd
    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 mrysh/qwen3.5-4b-prune_width_ffn15_lora_kd --local-dir ~/models/qwen3.5-4b-prune_width_ffn15_lora_kd
    # 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. Qwen3.5-4B-Basefamily root
  2. Qwen3.5-4B
  3. qwen3.5-4b-prune_width_ffn15
  4. qwen3.5-4b-prune_width_ffn15_lora_kd

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

mrysh/qwen3.5-4b-prune_width_ffn15_lora_kd
Author
mrysh
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