Skip to content
Other

Glimmer Scope SAEs

9 слоёв Muse Glimmer 30B под микроскопом

2 min read

Набор sparse autoencoders от Kroonen AI для Muse Glimmer 30B: по одному SAE на каждом третьем слое residual stream с 12 по 36. Каждый разбивает вычисления слоя на 65 536 признаков; это основа Glimmer Scope для разметки, steering и трассировки признаков между слоями.

Ключевые особенности

  • 9 SAE для слоёв 12, 15, 18, 21, 24, 27, 30, 33 и 36; вход — residual stream после decoder layer N (0-indexed из 52), d = 6 656.
  • В каждом слое 65 536 признаков, TopK с k = 64, decoder rows с unit norm; веса float32, 3.5 ГБ на слой.
  • Обучение: 70% FineWeb-Edu и 30% UltraChat 200k в чат-шаблоне Glimmer; normalized MSE плюс AuxK (k_aux = 512).
  • На 5M-токенном сэмпле web/chat срабатывают 519 383 из 589 824 признаков.
  • В релизе только SAE: без меток признаков, activation data и результатов steering; метки есть в atlas, но без тревожных тем.

Пример

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import json
REPO, LAYER = "kroonen-ai/glimmer-scope-saes", 24
cfg = json.load(open(hf_hub_download(REPO, f"layer-{LAYER}/cfg.json")))
sae = load_file(hf_hub_download(REPO, f"layer-{LAYER}/sae.safetensors"))
x = resid.float() * cfg["input_scale"]
vals, ids = ((x - sae["b_dec"]) @ sae["W_enc"] + sae["b_enc"]).topk(cfg["k"], -1)
ПодробнееVariance unexplained по слоям: 0.200–0.315; dead features: 0.5–1.7%.
Каждый SAE — float32, 3.5 ГБ; вход d = 6 656, TopK k = 64.
Нет feature labels, activation data и steering results; публичные метки — в atlas с фильтрацией тревожных тем.
Apache 2.0, как у Muse Glimmer 30B.
Discuss Telegram 2

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # kroonen-ai/glimmer-scope-saes
    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 kroonen-ai/glimmer-scope-saes --local-dir ~/models/glimmer-scope-saes
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to glimmer-scope-saes
  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

New posts about versions of glimmer-scope-saes (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

kroonen-ai/glimmer-scope-saes
Author
kroonen-ai
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

02

Ask AI about this model

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

Comments