Skip to content
Other

Subspace NLA: Decoding hidden states of Qwen3-8B

2 min read

The Subspace NLA (Natural Language Autoencoder) model is designed for AI model interpretability, allowing you to decompose Qwen3-8B layer activations into separate semantic components. The system works alongside a Block-Sparse Featurizer (BSF), which splits the 24th layer's activation into 16 active blocks within 64-dimensional subspaces. A verbalizer generates short text descriptions for each component, while a reconstructor assembles them back into a vector. This tool is useful for researchers and developers studying the internal decision-making mechanisms of LLMs (large language models).

Key features

  • Autoencoder construction: breaking down layer activations into independent subspaces
  • Verbalization: generating text descriptions for each active block
  • Vector reconstruction back into the model space using a trained projector
  • Fine-tuning via GRPO (Group Relative Policy Optimization) with a GRPO reward for projection accuracy

What you can do

  • Run text token analysis using the quick-start script for the model
  • Explore internal vector representations and semantic components of Qwen3-8B activations
  • Check vector reconstruction quality using the FVE (Fraction of Variance Explained) metric

Example

python scripts/hf_quickstart.py --nla asher577/subspace-nla-qwen3-8b --text "The capital of France is"
Read moreDetails
Benchmarks: FVE for the entire activation is 0.489 with a cosine similarity of 0.857
Requirements: runs on top of the base Qwen/Qwen3-8B model
Training: GRPO for 2,400 steps on a mixture of chat and web text
Limitations: requires the mandatory use of the paired BSF featurizer
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 15 GB · ready files17 GB
  • FP8 / INT8recommendedweights 7.6 GB8.9 GB
  • 4-bitweights 4.3 GB5.2 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 3080 · BF16 — from RTX 4090

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.

    # asher577/subspace-nla-qwen3-8b
    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 asher577/subspace-nla-qwen3-8b --local-dir ~/models/subspace-nla-qwen3-8b
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to subspace-nla-qwen3-8b
  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. Qwen3-8B-Basefamily root
  2. Qwen3-8B
  3. subspace-nla-qwen3-8b

New posts about versions of Qwen3-8B-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 4All on Hugging Face

About the model

asher577/subspace-nla-qwen3-8b
Author
asher577
Category
Other
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