Instructions to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16
# Run inference directly in the terminal:
llama cli -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16
# Run inference directly in the terminal:
llama cli -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16
# Run inference directly in the terminal:
./build/bin/llama cli -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Use Docker

docker model run hf.co/AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Llama new
LM Studio
Jan
Ollama
How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with Ollama:
```
ollama run hf.co/AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16
```
Unsloth Desktop

How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with Docker Model Runner:
```
docker model run hf.co/AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16
```

Lemonade

How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Run and chat with the model

lemonade run user.Qwen3.5-4B-AX8860-llama.cpp-BF16

List all available models

lemonade list

Hermes Agent

How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp:BF16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Qwen3.5-4B on AX8860 (llama.cpp + ggml-axera, 2026-11-05)

在爱芯 AX8860 的 NPU（4 核）上用 llama.cpp 运行 Qwen3.5-4B，支持文本和图像输入；16k 上下文；服务端可选用模型自带的 MTP 头做投机解码。 Qwen3.5-4B with text and image input on the 4-core NPU of the Axera AX8860, with llama.cpp; 16k context; the server can decode speculatively with the model's own MTP head.

文件 / Files

文件 / File	说明 / Description
`qwen3.5-4b-mxfp4.gguf`	模型权重（GGUF MXFP4，2.66 GB，md5 `e82f054c77152821eb8edd5ae572f2eb`）/ weights
`mmproj-qwen3.5-4b-bf16.gguf`	视觉编码器权重（BF16，676 MB，md5 `d64c1da66e6bd59ecbec5d57ccc13ec7`）/ vision encoder weights
`bin/`	llama.cpp 板端程序（静态链接，只依赖系统 `/opt/lib` 与 libc）/ the board programs (static; need only the system `/opt/lib` and libc)
`kernels/`	NPU kernel + `manifest.json`：prefill、16 层 decode group（两组）、输出头、3072-patch ViT、两个 IO 描述 kernel；投机解码用的 verify（4 行验证）、MTP 头（2 个）、草稿词表头 + `draft_vocab.i32` / the NPU kernels + `manifest.json`: prefill, a 16-layer decode group (two groups), the output head, the 3072-patch ViT, two IO descriptor kernels; for speculative decoding the verify (4 rows), the MTP head (2 kernels), the draft-vocabulary head + `draft_vocab.i32`
`run.sh`	运行脚本 / run script

权重只来自 GGUF：kernel 不含权重，加载时从 GGUF 转换到 NPU（约 9 s）；NPU 内存里只放一份权重（prefill 和 decode 共用）。 The weights come only from the GGUFs: the kernels hold none; they are converted from the GGUF at load (~9 s); one copy of the weights in NPU memory (shared by prefill and decode).

运行 / Run

环境 / Requirements：AX8860，SDK V0.5.0 或更新（cat /proc/ax_proc/version）/ SDK V0.5.0 or newer.

./run.sh "The capital of France is" 64                 # 文本补全 / text completion
./run.sh -i bus.jpg "Describe this image." 256         # 图像描述 / describe an image
./run.sh -c                                            # 命令行多轮对话 / multi-turn chat in the terminal
./run.sh -s 8080                                       # OpenAI 兼容服务 / OpenAI-compatible server
MTP=1 ./run.sh -s 8080                                 # 服务端 MTP 投机解码 / the server with MTP speculative decoding
curl http://<board>:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"你好"}]}'

直接调用程序时设置 / When calling the programs directly: LD_LIBRARY_PATH=/opt/lib GGML_AXERA_KERNEL_DIR=<repo>/kernels，-ngl 99，-c 不超过 16384，服务端 -np 1；投机解码（只有 llama-server 支持）--spec-type draft-mtp --spec-draft-n-max 3，不加则为普通解码 / speculative decoding (llama-server only) --spec-type draft-mtp --spec-draft-n-max 3; without it: plain decoding。GGML_AXERA_PROF=1 退出时打印各阶段耗时 / prints per-stage times at exit。

性能 / Performance

AX8860，SDK V0.5.1，NPU 4 核 / 4 cores，贪心解码 / greedy decoding：

项 / Item	值 / Value
加载后 NPU 内存（CMM）/ NPU memory after load	3.4 GB
加载后系统内存 / host memory after load	1.0 GB
加载 / load	9 s（服务端 / server 21 s）
4k prompt prefill（4008 token）	7.7k tokens/s（每 256 token 页 32.6 ms / 32.6 ms per 256-token page）
decode（短上下文 / short context）	60 tokens/s（服务端长回答 / server, long answers 53 - 55）
图像编码 / image encoding（3072 patch = 768 图像 token / image tokens）	0.20 s

MTP 投机解码（服务端，贪心，同一组请求，tokens/s）/ MTP speculative decoding (server, greedy, the same requests, tokens/s)：

回答 / Answer	普通 / plain	MTP
数学推理（700 token）/ maths reasoning	54.5	78.5
代码（700 token）/ code	55.0	69.1
翻译 / translation	54.0	65.1
事实列表（700 token）/ a list of facts	53.7	64.6
中文讲解（800 token）/ a Chinese explanation	53.8	61.6
英文作文（800 token）/ an English essay	53.9	55.9
英文小说（1500 token）/ an English story	53.6	53.7
中文作文 / a Chinese essay	52.9	48.8

收益取决于草稿的接受率：结构化的内容（代码、推理、列表）快 15 - 45%，自由写作基本不变，所以默认不开 / the gain follows the draft acceptance: 15 - 45% on structured text (code, reasoning, lists), none on free prose, hence off by default.

正确性 / Correctness

"The capital of France is" → "Paris. …"；图像描述正确（车辆、文字、人物）/ image descriptions correct.
中文 PPL（内部测试集，-c 512 --chunks 3）：prefill 30.73，逐 token decode（-b 1）30.56，投机解码的 4 行验证（-ub 4）30.53；同一 GGUF 在 CPU 上 30.10 / Chinese PPL on an internal set: 30.73 through the prefill, 30.56 through the token-by-token decode (-b 1), 30.53 through the 4-row verification (-ub 4); the same GGUF on the CPU: 30.10.
本版修正了上一版 decode 的一个错误（manifest.json 把 decode 包里 delta-net 的两组缩放标成了 MXFP8，decode 的 PPL 为 48.7，长回答容易重复）；kernel 文件不变 / this release fixes a decode error of the previous one (manifest.json tagged two delta-net scale sets of the decode packs as MXFP8: decode PPL 48.7, long answers prone to repetition); the kernel files are unchanged.
投机解码的输出由目标模型逐个验证，与普通解码质量一致（两者累加顺序不同，长回答可能在近似平局处分叉）/ every speculative token is verified by the target model: the quality equals plain decoding (the two accumulate in a different order, so long answers may diverge at near-ties).
同一输入多次运行，结果偶有极小差异（NPU 上累加顺序不固定），不影响输出质量。 Repeated runs of one input may differ slightly (the NPU's accumulation order varies); the output quality is the same.

已知限制 / Known limits

运行时独占 NPU 4 核，同时运行其他 NPU 程序会出错 / takes all 4 NPU cores; do not run other NPU programs alongside.
单序列：服务端须 -np 1 / one sequence: the server needs -np 1.
上下文最大 16384 / the context is at most 16384.

源码 / Sources

llama.cpp：ZHEQIUSHUI/llama.cpp axera-dev @ 57a30bd（ggml/src/ggml-axera/models/qwen35_whole）
kernel：ZHEQIUSHUI/ax-dsl-kernels fix/qwen35-4b-manifest @ 4822156（models/qwen35_4b/whole，qwen35_4b.manifest.json，kernels.md5）

Downloads last month: 178

GGUF

Model size

4B params

Architecture

qwen35

Hardware compatibility

4-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including AXERA-TECH/Qwen3.5-4B-AX8860-llama.cpp

AX8860

Collection

14 items • Updated 1 day ago