Instructions to use Jab1718/qwen3.8-flash-coder-26gb-gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Jab1718/qwen3.8-flash-coder-26gb-gguf 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 Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Use Docker

docker model run hf.co/Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

LM Studio
Jan

vLLM

How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Jab1718/qwen3.8-flash-coder-26gb-gguf"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Jab1718/qwen3.8-flash-coder-26gb-gguf",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Ollama
How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with Ollama:
```
ollama run hf.co/Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M
```
Unsloth Desktop

How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

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": "Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with Docker Model Runner:
```
docker model run hf.co/Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M
```

Lemonade

How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Run and chat with the model

lemonade run user.qwen3.8-flash-coder-26gb-gguf-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

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 Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Jab1718/qwen3.8-flash-coder-26gb-gguf with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M

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 "Jab1718/qwen3.8-flash-coder-26gb-gguf:Q4_K_M" \
  --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"

⚡ Qwen3.8-Flash-Coder-26GB-GGUF (160 Experts Hardware-Aligned Coding Subnet)

This repository provides GGUF quantized formats of Jab1718/qwen3.8-flash-coder-85gb-bf16, a 160-expert coding subnet sliced from Qwen3.8-Flash-Next (335GB) using the moe-slice toolkit.

📦 Provided GGUF Files

File Name	Quantization	Size on Disk	Target Hardware	Recommended Use Case
`qwen3.8-flash-coder-26gb-q4_k_m.gguf`	Q4_K_M	26.43 GB (27 GB)	1x 24GB GPU (RTX 3090/4090), RTX 5000 Ada (32GB), or 32GB Mac	🔥 92% VRAM reduction from 335GB base! Run locally on single consumer GPU!
`qwen3.8-flash-coder-43gb-q8_0.gguf`	Q8_0	42.25 GB (43 GB)	2x 24GB GPUs / 1x 48GB GPU / 64GB+ RAM Mac	Maximum reasoning fidelity (~99.5% BF16)
`Modelfile`	Text config	<1 KB	Ollama	Quick import and deployment in Ollama

🚀 Quick Start with Ollama

Step 1: Download the GGUF file and Modelfile

huggingface-cli download Jab1718/qwen3.8-flash-coder-26gb-gguf qwen3.8-flash-coder-26gb-q4_k_m.gguf --local-dir .
huggingface-cli download Jab1718/qwen3.8-flash-coder-26gb-gguf Modelfile --local-dir .

Step 2: Create and run model in Ollama

ollama create qwen3.8-flash-coder -f Modelfile
ollama run qwen3.8-flash-coder "Implement a thread-safe LRU Cache with TTL expiration in Rust."

🍏 Apple Silicon & Mac Setup Notes (Metal Acceleration)

Runtime Compatibility Notice for macOS Users:

Ollama & Legacy Runtimes: Bundled Ollama versions (e.g., v0.32.5 or earlier) and older DS4 binaries hardcode expert_count=512 and will reject this file (expected expert_count=512 … got 160).

Recommended Runtime: Build and run using the latest llama.cpp master with Metal support (which includes native compatibility PRs for arbitrary MoE expert counts like qwen4exp).

⚡ Verified Apple Silicon Performance Profile (Tested on Apple M5 Pro, 64 GB Unified Memory, macOS 26)

Extraordinary Memory Efficiency: Loads at full 64K context (-c 65536) using only ~33 GB wired RAM (~46% RAM free) — massive headroom for concurrent developer workflows on 64 GB Macs!
Blazing Fast Decode: ~32 tokens/second sustained decode throughput.
Near-Instant Initialization: Cold load in ~10 seconds via memory-mapped I/O.

Recommended `llama-server` Launch Command:

./llama-server \
  -m ./qwen3.8-flash-coder-26gb-q4_k_m.gguf \
  -c 65536 \
  -ngl 999 \
  --host 0.0.0.0 \
  --port 8080

💡 Prompting Guidelines & Best Practices

1. Code Review & Bug-Finding Prompts (Preventing Empty Content)

The Symptom: In analytical/code-auditing tasks (e.g., "Find all bugs in this function"), certain frontends/chat templates may capture reasoning inside <think>...</think> but route near-empty text (0–2 chars) into the final content field upon encountering finish_reason=stop.
Root Cause: Chat template stop-token routing or missing explicit conclusion markers outside the thought block.
Solution (Explicit Output Routing): Instruct the model explicitly to provide its structured audit report outside the thought tags:

# Recommended Prompt Pattern:
Review the following code for concurrency bugs and memory leaks.
Think thoroughly through edge cases inside <think>...</think>.
CRITICAL: After closing the </think> tag, you MUST provide your final structured bug report (specifying Bug Name, Line Number, Severity, and Recommended Fix) directly in your response content.

2. Surgical Code Refactoring (Preventing Over-Rewrites)

To prevent the model from rewriting entire algorithms from scratch (which can introduce edge-case regressions like off-by-one errors):
Best Practice: Prompt the model for a minimal diff or surgical patch:

# Recommended Refactoring Prompt:
Refactor the following function to fix the infinite loop bug.
Provide ONLY a minimal surgical unified diff patch (+/- lines) without rewriting unaffected logic.

3. Domain Focus & General Math Trade-off

Note on Math/GSM8K: This 160-expert subnet was intentionally extracted by prioritizing Systems Programming, Algorithmic DP, and Tool-Calling trajectories. General word-problem arithmetic (GSM8K) was pruned by design to give maximum capacity to multi-language software development.

📊 Benchmark Capabilities

Rust (Systems): 10/10 (100.0%) Pass@1
C++20 (Modern Systems): 10/10 (100.0%) Pass@1
TypeScript (Fullstack): 4/5 (80.0%) Pass@1
Coding Agent (Diff/FIM/Bugfix): 16/20 (80.0%) Pass@1
Python Algorithms: 39/50 (78.0%) Pass@1

📜 Citation

@software{thainq2026moeslice_gguf,
  author = {Thai Nguyen},
  title = {Qwen3.8-Flash-Coder-GGUF: Hardware-Aligned Coding Subnet for llama.cpp and Ollama},
  url = {https://github.com/Jab1718/Moe-slices},
  year = {2026}
}

Downloads last month: 7,016

GGUF

Model size

43B params

Architecture

qwen4exp

Hardware compatibility

4-bit

8-bit

Model tree for Jab1718/qwen3.8-flash-coder-26gb-gguf

Base model

Qwen/Qwen3.8-Flash-Next

Finetuned

Jab1718/qwen3.8-flash-coder-85gb-bf16

Quantized

(3)

this model

⚡ Qwen3.8-Flash-Coder-26GB-GGUF (160 Experts Hardware-Aligned Coding Subnet)

📦 Provided GGUF Files

🚀 Quick Start with Ollama

Step 1: Download the GGUF file and Modelfile

Step 2: Create and run model in Ollama

🍏 Apple Silicon & Mac Setup Notes (Metal Acceleration)

⚡ Verified Apple Silicon Performance Profile (Tested on Apple M5 Pro, 64 GB Unified Memory, macOS 26)

Recommended llama-server Launch Command:

💡 Prompting Guidelines & Best Practices

1. Code Review & Bug-Finding Prompts (Preventing Empty Content)

2. Surgical Code Refactoring (Preventing Over-Rewrites)

3. Domain Focus & General Math Trade-off

📊 Benchmark Capabilities

📜 Citation

Model tree for Jab1718/qwen3.8-flash-coder-26gb-gguf

Recommended `llama-server` Launch Command: