Skip to content
Other

Qwen3.5-0.8B на AX8860

мультимодальная модель на NPU через llama.cpp

2 min read

AXERA-TECH выложила сборку Qwen3.5-0.8B для 4-ядерного NPU Axera AX8860 на llama.cpp и ggml-axera. Модель принимает текст и изображения; веса идут в GGUF Q8_0 (0.83 GB), визуальный энкодер — отдельный BF16 GGUF (0.21 GB). Контекст — до 16384 токенов.

Ключевые особенности

  • Поддержка текста и изображений: есть ViT-энкодер и режим описания картинки.
  • На AX8860: prefill 4k промпта — 21600 tokens/s, decode — 125.6 tokens/s, кодирование изображения — 80 ms + 20 ms CPU.
  • Веса Q8_0 при загрузке конвертируются в MXFP8 (E4M3 + E8M0 scale на каждые 32), в NPU-памяти хранится одна копия.
  • Prefill — один kernel на всю модель, decode — по одному kernel на 12 слоёв; загрузка весов из GGUF занимает 6.9 s.
  • run.sh умеет текстовое дополнение, описание изображения, чат в терминале и OpenAI-совместимый сервер.
  • Ограничения: одна последовательность, контекст до 16384, NPU 4 ядра занимаются эксклюзивно, спекулятивного декодирования нет.

Пример

./run.sh "The capital of France is" 64
./run.sh -i bus.jpg "Describe this image." 256
./run.sh -s 8080
ПодробнееКитайский PPL: 40.76 на NPU против 40.13 на CPU (+1.6%).
Память: 1.26 GB NPU в пике и 0.47 GB host; загрузка — 8.3 s, конвертация весов — 6.9 s.
Требуется AX8860 и SDK V0.5.0+; на сервере нужен -np 1.
Discuss Telegram 1 263

Will it run?

Runs in GGUF Q8_0

The most precise of the ready quantised files fits.

  • GGUF Q8_0recommendedweights 0.8 GB · ready files1.4 GB

Tip: go 8-bit.GGUF Q8_0 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: GGUF Q8_0 — 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.

    # AXERA-TECH/Qwen3.5-0.8B-AX8860-llama.cpp
    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 AXERA-TECH/Qwen3.5-0.8B-AX8860-llama.cpp --local-dir ~/models/Qwen3.5-0.8B-AX8860-llama.cpp
    # 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-AX8860-llama.cpp

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

AXERA-TECH/Qwen3.5-0.8B-AX8860-llama.cpp
Author
AXERA-TECH
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

2631

Ask AI about this model

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

Comments