Instructions to use AXERA-TECH/Qwen3.5-9B-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/Qwen3.5-9B-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/Qwen3.5-9B-AX8860-llama.cpp:BF16
# Run inference directly in the terminal:
llama cli -hf AXERA-TECH/Qwen3.5-9B-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/Qwen3.5-9B-AX8860-llama.cpp:BF16
# Run inference directly in the terminal:
llama cli -hf AXERA-TECH/Qwen3.5-9B-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/Qwen3.5-9B-AX8860-llama.cpp:BF16
# Run inference directly in the terminal:
./build/bin/llama cli -hf AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp:BF16

Use Docker

docker model run hf.co/AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp:BF16

Llama new
LM Studio
Jan
Ollama
How to use AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp with Ollama:
```
ollama run hf.co/AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp:BF16
```
Unsloth Desktop

How to use AXERA-TECH/Qwen3.5-9B-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/Qwen3.5-9B-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/Qwen3.5-9B-AX8860-llama.cpp:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp with Docker Model Runner:
```
docker model run hf.co/AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp:BF16
```

Lemonade

How to use AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull AXERA-TECH/Qwen3.5-9B-AX8860-llama.cpp:BF16

Run and chat with the model

lemonade run user.Qwen3.5-9B-AX8860-llama.cpp-BF16

List all available models

lemonade list

Hermes Agent

How to use AXERA-TECH/Qwen3.5-9B-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/Qwen3.5-9B-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/Qwen3.5-9B-AX8860-llama.cpp:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AXERA-TECH/Qwen3.5-9B-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/Qwen3.5-9B-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/Qwen3.5-9B-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.

Qwen3.5-9B on AX8860 (llama.cpp + ggml-axera, 2026-11-04)

在爱芯 AX8860 的 NPU（4 核）上用 llama.cpp 运行 Qwen3.5-9B，支持文本和图像输入；三套 kernel 对应 16k / 128k / 256k 上下文。 Qwen3.5-9B with text and image input on the 4-core NPU of the Axera AX8860, with llama.cpp; three kernel sets for 16k / 128k / 256k contexts.

文件 / Files

文件 / File	说明 / Description
`gguf/qwen3.5-9b-mxfp4.gguf`	模型权重（GGUF MXFP4，6.00 GB，md5 `57b4b4b8fe3e7d05af29019438e5369a`）/ weights
`gguf/mmproj-qwen3.5-9b-bf16.gguf`	视觉编码器权重（BF16，921 MB，md5 `69d1a645084d0f31d4296a82dce991cf`）/ vision encoder weights
`bin/`	llama.cpp 板端程序与共享库（`run.sh` 设置 `LD_LIBRARY_PATH`；还需系统 `/opt/lib` 与 libc）/ the board programs and shared libraries
`kernels/`	默认 / default：128k 上下文的 NPU kernel + `manifest.json` / the 128k-context kernels
`kernels_16k/`	16k 上下文：NPU 内存最省，短对话 / agent 场景 / 16k context: the smallest NPU memory
`kernels_256k/`	256k 上下文（模型原生最大）/ 256k context (the model's native maximum)
`run.sh`	运行脚本 / run script

每套 kernel：prefill、16 层 decode group（两组）、输出头、3072-patch ViT、两个 IO 描述 kernel + manifest.json；三套只有 prefill 和 decode group 的上下文不同。 Each set: prefill, a 16-layer decode group (two groups), the output head, the 3072-patch ViT, two IO descriptor kernels + manifest.json; only the prefill and decode group change with the context.

权重只来自 GGUF：kernel 不含权重，加载时从 GGUF 转换到 NPU；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; 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 assets/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
KERNELS=16k ./run.sh -s 8080                           # 16k 套件（省内存）/ the 16k set
KERNELS=256k CTX=200000 ./run.sh -s 8080               # 256k 套件，llama.cpp 上下文 200000
curl http://<board>:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"你好"}]}'

KERNELS（128k 默认 / 16k / 256k）选 kernel 套件；CTX 是传给 llama.cpp 的 -c，默认等于套件的上下文，可以更小。 KERNELS (128k, the default / 16k / 256k) picks the kernel set; CTX is llama.cpp's -c, the set's context by default, may be smaller.

直接调用程序时设置 / When calling the programs directly: LD_LIBRARY_PATH=/opt/lib:$PWD/bin GGML_AXERA_KERNEL_DIR=<repo>/kernels （或 kernels_16k / kernels_256k），-ngl 99，-c 不超过套件的上下文，服务端 -np 1。 GGML_AXERA_PROF=1 退出时打印各阶段耗时；GGML_AXERA_TEMP_MAX=<°C>（默认 100，0 关闭）在 NPU 温度过高时暂停 kernel。

性能 / Performance

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

项 / Item	`kernels_16k`	`kernels`（128k）	`kernels_256k`
加载后 NPU 内存（CMM）/ NPU memory after load	6.27 GB	10.18 GB	14.61 GB
加载后系统内存 VmHWM / host memory after load	1.31 GiB	1.32 GiB	1.32 GiB
加载 / load	30 s	33 s	35 s
4k prompt 的 prefill 页耗时 / per-page prefill	57.1 ms（16 页 / pages，913.8 ms）	-	-
短上下文 decode / short-context decode	42.7 tokens/s	-	-
32k prompt prefill / decode	-	2434 tokens/s / 30.5 tokens/s	2458 tokens/s / 30.3 tokens/s
64k prompt prefill / decode	-	2028 tokens/s / 23.6 tokens/s	1990 tokens/s / 23.3 tokens/s
120k prompt（113,463 token）prefill / decode	-	1529 tokens/s / 17.9 tokens/s	1481 tokens/s / 17.7 tokens/s
英文小说 PPL（`-c 4096`）/ long-text PPL	-	5.0061	5.0139

所有长上下文 decode 计时均为 23 次。长上下文下 attention 要读整个 KV，decode 随 -c 变慢。 All long-context decode timings have 23 runs; attention reads the whole KV, so decode slows with context.

正确性 / Correctness

16k 贪心文本 / 16k greedy text：rc 0，md5 a7785d21b97952f6e180618317c8c836；输出连贯，补充的 -n 256 运行回答 “Paris”。
16k 中文 PPL（内部测试集，-c 512）/ Chinese PPL: NPU 20.8258 +/- 2.88086，同 GGUF 的 host CPU 20.5512 +/- 2.89327（+1.34%，门限 5%）/ same-GGUF host CPU control.
16k 图像 / 16k image：rc 0，与旧 host 逐字节一致，md5 4b8ad6e3a712219b00f65917be7ad658；内容正确 / coherent.
大海捞针 / needle in a haystack（暗号 7391-ALPHA-BRAVO 藏在英文小说 30% 处 / code at 30% of an English novel）：128k 套件 32k / 64k / 120k 全部命中；256k 套件 32k / 64k / 120k 全部命中。
三套 kernel 的 prefill / quantized KV 索引范围检查通过；真实 32k / 64k / 120k prompt 分别为 30,917 / 61,177 / 113,463 token。128k prefill 四段 8 层 call 的链式 cos 为 mean 0.999894 / min 0.999134；256k 八段 4 层 call 为 mean 0.999894 / min 0.999347。
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.
板子 16 GB 系统内存：llama-perplexity 的 -c 不超过 4096（它把整段上下文的 logits 放在系统内存里）/ with the board's 16 GB host memory, llama-perplexity works up to -c 4096.
上下文随 -c 增大，decode 变慢（attention 要读整个 KV）：64k 上下文 23.3 tokens/s，113k 上下文 17.7 tokens/s / decode slows with the context.
长 prefill 会加热 NPU；host 在温度超过 GGML_AXERA_TEMP_MAX（默认 100 °C）时暂停 kernel，这是保护，不是错误 / long prefill heats the NPU; the host pauses above the limit.

源码 / Sources

llama.cpp：ZHEQIUSHUI/llama.cpp axera-dev @ 20a3e30 （ggml/src/ggml-axera/models/qwen35_whole）
kernel：ZHEQIUSHUI/ax-dsl-kernels main @ a6064b5 （models/qwen35_9b/whole，kernels.md5 列出三套 / lists the three sets）

Downloads last month: 7

GGUF

Model size

9B 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/Qwen3.5-9B-AX8860-llama.cpp

AX8860

Collection

13 items • Updated about 10 hours ago