Skip to content
Computer vision

Qwen3.8-27B-V13-b1

27B-LLM с эпизодической памятью

1 min read

Экспериментальная 27,78B-модель от LTM-Lab: базовые веса DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored без изменений, но с архитектурой qwen_dense_v13. Это Qwen3.5 dense с 64 слоями, vision tower, эпизодической KV-памятью и latent foresight. Код архитектуры поставляется с моделью — нужен trust_remote_code=True.

Ключевые особенности

  • Архитектура qwen_dense_v13: 48 слоёв Gated DeltaNet linear-attention и 16 full-attention слоёв.
  • Эпизодическая KV-память: записи сохраняются в SQLite-хранилище и вспоминаются в следующих вызовах; есть read-only режим, статистика и forget_episode.
  • Latent foresight: 16 предикторов примерно на 0,1B параметров, по одному на full-attention слой; рантайм дообучает их после 256 переходов и при каждом удвоении.
  • MTP-голова для speculative decoding и chat template с reasoning_effort low/medium/xhigh.
  • В тесте recall — 0/40: хранилище читается, но записанные значения не возвращаются.
  • Нужны Transformers 5.x и flash-linear-attention; хранилище памяти держите на локальном диске.

Пример

import os, torch
from transformers import AutoModelForImageTextToText, AutoTokenizer
os.environ["QWEN_DENSE_V13_MEMORY_HOME"] = "/path/v13-memory"
repo = "LTM-Lab/Qwen3.8-27B-V13-b1"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda").eval()
print(model.memory_stats())
ПодробнееБенчмарки: recall 0/40, generation 6/6 без лишних <think> и циклов, held-out NLL 0.545.
Тестировали на одной GPU 96 GB: ~13 ток/с с памятью и ~52 ток/с без неё.
Ограничения: похожие факты могут путаться, база uncensored, стандартные бенчмарки не заявлены.
Discuss Telegram 0 41

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 52 GB · ready files57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 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.

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.

    # LTM-Lab/Qwen3.8-27B-V13-b1
    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 LTM-Lab/Qwen3.8-27B-V13-b1 --local-dir ~/models/Qwen3.8-27B-V13-b1
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.8-27B-V13-b1
  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. Qwen3.8-27Bfamily root
  2. Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
  3. Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored
  4. Qwen3.8-27B-V13-b1

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

LTM-Lab/Qwen3.8-27B-V13-b1
Author
LTM-Lab
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

410

Ask AI about this model

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

Comments