Instructions to use unsloth/Qwen3.8-Flash-Next-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 unsloth/Qwen3.8-Flash-Next-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 unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

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 unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Llama new
LM Studio
Jan

vLLM

How to use unsloth/Qwen3.8-Flash-Next-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/Qwen3.8-Flash-Next-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": "unsloth/Qwen3.8-Flash-Next-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Ollama
How to use unsloth/Qwen3.8-Flash-Next-GGUF with Ollama:
```
ollama run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
```
Unsloth Desktop

How to use unsloth/Qwen3.8-Flash-Next-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

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": "unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/Qwen3.8-Flash-Next-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/Qwen3.8-Flash-Next-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.Qwen3.8-Flash-Next-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use unsloth/Qwen3.8-Flash-Next-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 unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

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 unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/Qwen3.8-Flash-Next-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

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 "unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL" \
  --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"

UD-IQ4_XS: canned, off-topic "graph cycle proof" reasoning replaces real thinking on agentic tool-call turns (~100K context)

#101

by Uer9 - opened 1 day ago

Discussion

Uer9

1 day ago

Summary
During a long agentic session (Hermes Agent, llama-server, ~100K tokens of context), 8 of 74 assistant turns had their entire reasoning block replaced by the same unrelated passage: a plan for proving that every graph with n vertices and more than n−1 edges contains a cycle. The wording varied slightly each time but was otherwise the same. Each of those turns still ended with a sensible, on-task tool call, so the task finished correctly. The reasoning text had nothing to do with the conversation, which was a server repair task with no math in it.

Setup

Model: UD-IQ4_XS (4 shards) + mmproj-F16.gguf; MTP draft mtp-Qwen3.8-Flash-Next-Q8_0.gguf (--spec-type draft-mtp --spec-draft-n-max 3)
llama.cpp: built from source, commit 7d7f028 (2026-11-08), Metal
Hardware: Mac Studio M4 Max, 128 GB, macOS 27.0.1
Server flags: -c 262144 -np 1 --jinja --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0 (thinking-mode settings from the model card)
Client: Hermes Agent over an OpenAI-compatible /v1/chat/completions endpoint, with streaming and tool calling. The agent is set to reasoning_effort: medium, and preserve_thinking was left at the template default.
A single slot (-np 1), so no concurrent requests shared the slot. The server log shows only this session, plus a few tiny 16-token requests between turns.

What happened

8 of the 74 assistant turns in one session were affected. They clustered over about an hour, with roughly 100K–104K tokens in the slot.
In every affected turn, the reasoning started with "The user is asking for a mathematical proof." The rest was ~800 characters planning a proof by induction or contrapositive, ending with "Write the proof concisely." After that came a correct, on-task tool call.
In one turn the reasoning read "Two methods are needed. This is a well-known standard result, so I'll provide two concise proofs. No tools needed." That turn still ended with a tool call.
The other 66 turns in the same session had normal, on-topic reasoning.

Example (one affected turn, reasoning field only, lightly trimmed):

The user is asking for a mathematical proof. The proposition to be proved: every graph with n vertices and more than n−1 edges (i.e., at least n edges) contains a cycle. Equivalent proposition: an acyclic graph (forest) has at most n−1 edges. Proof by induction or by contrapositive. … Proof: take a maximal path. Its endpoint has degree 1. … Write the proof concisely.

What I ruled out

Cross-talk from another client: there's one slot, and no other client sent anything resembling this prompt. The canned reasoning and the on-task tool call came back in the same response.
A leftover test or benchmark: no test prompt anywhere on the system contains this text.

Questions

Has anyone else seen stereotyped or canned reasoning like this, especially at long context or on tool-call turns?
Is this more likely an IQ4_XS quantization artifact, or does it also happen on Q4_K_XL / Q8_0 / BF16? (On 128 GB I can't easily test the bigger quants at full context.)
Is a different reasoning_effort or preserve_thinking setting recommended for long agentic sessions?

I can provide the full reasoning text for all 8 affected turns if that helps.

Uer9

1 day ago

Update: likely mitigation found.

I set preserve_thinking to false server-side (llama-server --chat-template-kwargs '{"preserve_thinking":false}', everything else unchanged: same UD-IQ4_XS file, llama.cpp build, sampling, 262K context, MTP). I confirmed via /apply-template that prior turns' reasoning is no longer re-sent; reasoning within the current turn is still kept, and tool-call loops work normally.

Since then: two long Hermes Agent sessions today (one still running) with no occurrence of the canned graph-cycle passage. Before the change it showed up in 8 of 74 assistant turns in a single ~100K session.

That's a small sample, but it fits a separate report of Flash Next echoing its own earlier reasoning more often as more copies accumulate in context, also traced to preserve_thinking: true. It suggests this is a base-model/preserved-thinking feedback effect rather than an IQ4_XS quantization artifact.

For anyone hitting this in long agentic sessions: try preserve_thinking: false first. I'll post again if it comes back.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment