Instructions to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-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 ji-farthing/Qwen3.8-Flash-Next-ik-llama-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 ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF
# Run inference directly in the terminal:
llama cli -hf ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

Install from WinGet (Windows)

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

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 ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF
# Run inference directly in the terminal:
./llama-cli -hf ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

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 ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

Use Docker

docker model run hf.co/ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

LM Studio
Jan

vLLM

How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ji-farthing/Qwen3.8-Flash-Next-ik-llama-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": "ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

Ollama
How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF with Ollama:
```
ollama run hf.co/ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF
```
Unsloth Desktop

How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

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": "ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF with Docker Model Runner:
```
docker model run hf.co/ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF
```

Lemonade

How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

Run and chat with the model

lemonade run user.Qwen3.8-Flash-Next-ik-llama-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-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 ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

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 ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

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 "ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF" \
  --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"

HuggingFace's download widget (right pane -->) does not recognize _KT ggufs. Download links are in the table below.

Qwen3.8-Flash-Next, ik_llama.cpp GGUF

Text-only GGUF conversions of Qwen3.8-Flash-Next, quantized with an importance matrix calibrated for agentic coding, plus the model's own MTP draft head as a companion file for speculative decoding.

All three files were built directly from Qwen's BF16 release.

Files

File	Quant	Download	Experts	PLE table	Dense
`Qwen3.8-Flash-Next-ik_llama-IQ3_KT-00001-of-00002.gguf`	IQ3_KT	41.70 GiB
`Qwen3.8-Flash-Next-ik_llama-IQ3_KT-00002-of-00002.gguf`	IQ3_KT	38.52 GiB
IQ3_KT total		80.22 GiB	50.39 GiB	26.84 GiB	2.86 GiB
`Qwen3.8-Flash-Next-ik_llama-IQ4_KT-00001-of-00002.gguf`	IQ4_KT	44.57 GiB
`Qwen3.8-Flash-Next-ik_llama-IQ4_KT-00002-of-00002.gguf`	IQ4_KT	43.85 GiB
IQ4_KT total		88.42 GiB	58.59 GiB	26.84 GiB	2.86 GiB
`Qwen3.8-Flash-Next-MTP-ik_llama-IQ4_KT.gguf`	IQ4_KT	2.21 GiB
`qwen4exp-agentic-v3.imatrix`	n/a	312 MiB
`qwen4exp-mtp-v1-draft.imatrix`	n/a	6.6 MiB

The two targets differ only in the routed experts: ffn_gate_exps and ffn_up_exps are IQ3_KT in one file and IQ4_KT in the other. Everything else is the same in both. ffn_down_exps has a row length of 640, which no 256-block type accepts, so it is IQ4_NL in both files. The per-layer token embedding (PLE) table is 26.8 GiB and is IQ4_NL for the same reason. Attention, SSM, hyper-connection and shared-expert tensors are IQ4_KT. token_embd is Q8_0.

MTP draft head

Qwen3.8-Flash-Next-MTP-ik_llama-IQ4_KT.gguf is the model's own next-token prediction head, converted from the same BF16 revision as the targets. It is not a standalone model. Load it beside either target with --model-draft or -md and enable it with --spec-type mtp:....

The head carries its own hyper-connection mixer, so the loader binds that rather than falling back to the trunk's. Its importance matrix was collected paired against the IQ3_KT target, on the same corpus as the targets, so the head is calibrated for the activations the quantized target actually produces.

GGML_CUDA_NO_PINNED=1 llama-server \
   -m Qwen3.8-Flash-Next-ik_llama-IQ3_KT-00001-of-00002.gguf \
   --model-draft Qwen3.8-Flash-Next-MTP-ik_llama-IQ4_KT.gguf \
   --spec-type mtp:n_max=1,p_min=0.0 \
   --defer-ple -ngl 49 -ot 'ffn_.*_exps=CPU' \
   -c 131072 -b 2048 -ub 2048 -fa on -ctk q8_0 -ctv q8_0 -ictk q8_0 --jinja

Validated with RTX 4070 with 12 GB VRAM and 64 GB system RAM: every layer on the GPU except the routed experts, which stay on host, and the PLE table left on disk. This serves 128K context in 6.7 GiB of VRAM at about 20 tokens per second.

Runtime

These files need ik_llama.cpp. The KT trellis types exist only there. The targets load on current main. The draft head needs a build that includes qwen4exp MTP support (PR #2369).

--defer-ple matters on any machine with less RAM than the file. The PLE table is 26.8 GiB and is read a few rows per token, so it can stay on disk. Without the flag the loader pulls it resident. The routed experts are read every token and do need to fit in RAM or VRAM: 50.4 GiB for IQ3_KT, 58.6 GiB for IQ4_KT.

Enabling the MTP head allocates a second compute buffer the same size as the main one.

Validation

Measured on one RTX 4070, 64 GB RAM, Core i7-11700K.

Perplexity and KL divergence against the BF16 source

Reference is the BF16 GGUF converted from the released weights, the same file both targets were quantized from. Both corpora are held out and share no text with the calibration corpus. The code corpus is CPython and PostgreSQL source. 10 chunks of 2048 tokens per corpus, one sequence per batch, every file measured with the same binary and the same flags.

File	English PPL	vs BF16	mean KLD	same top-1	Code PPL	vs BF16	mean KLD	same top-1
BF16 (reference)	3.3316	n/a	n/a	n/a	1.7218	n/a	n/a	n/a
IQ3_KT	3.6119	+8.4%	0.151	85.1%	1.8273	+6.1%	0.135	92.1%
IQ4_KT	3.5625	+6.9%	0.133	85.7%	1.8209	+5.8%	0.125	92.4%

Same top-1 is the share of positions where the quantized file's most likely token is the reference's. On code the median KL divergence is 0.003 for both files, so the typical code token is reproduced almost exactly and the mean is carried by a thin tail. The IQ4_KT buys a 12.5% lower mean KLD on English and 7.6% on code for 8.2 GiB more file.

Calibration

The importance matrix was computed over 700 chunks of 2048 tokens. The corpus was built to match an agentic coding deployment rather than general capability: 45% agentic and technical English, 30% code (including the entirety of ik_llama and llama.cpp source code), 15% general English prose, 10% Chinese prose, measured in tokens. The Chinese share is there so the quantizer does not treat those channels as unimportant; it is not meant to tune the model toward Chinese.

Conversion notes

Source: Qwen/Qwen3.8-Flash-Next at revision de4b8e4d43b917e7706784d8bb445c9af86a3540.

Targets converted with convert_hf_to_gguf.py from llama.cpp at commit f8dbcd618, unmodified, --outtype bf16. The converter drops the vision tower and the MTP head. Quantized with llama-quantize from ik_llama.cpp at commit 15dddc60b, with a chunked dequantization change so the 51.2G-element PLE table fits a 64 GB machine; that change was checked identical against the unmodified quantizer on tensors that fit both ways.

The head was converted from the same revision with a local converter patch that exports the mtp.* tensors, which the stock converter omits by design, and quantized with the same recipe discipline as the targets.

Recipe, shared by both targets except for the two routed-expert rules:

ffn_gate_exps=iq3_kt   (iq4_kt in the IQ4_KT file)
ffn_up_exps=iq3_kt     (iq4_kt in the IQ4_KT file)
ffn_down_exps=iq4_nl
per_layer_token_embd.weight=iq4_nl
ple_conv1d=f16
token_embd.weight=q8_0
everything else=iq4_kt

sha256 of the unsplit files, and of the shards as uploaded:

546caa277e273708317d216c58b8dfe35352182b5ff63c33d24414082b3c8660  qwen4exp-agentic-v3.imatrix
366ca93c5f99308d431c999c7f487b9e225741b2c6d7666db159a1b860c3700e  Qwen3.8-Flash-Next-ik_llama-IQ3_KT.gguf
81a227c0732739724323b6fc4d36a887ee1ef4e4b0635d8ac8aedf3992c28333  Qwen3.8-Flash-Next-ik_llama-IQ4_KT.gguf
b15000cd9618129885c098c3fd0859a8897be984d9f0f883a183c07827710895  Qwen3.8-Flash-Next-MTP-ik_llama-IQ4_KT.gguf
2ec71b5d7cd03f2ca43ddb722b7178b240ba5658f5041e232ee01782e15b56d0  Qwen3.8-Flash-Next-ik_llama-IQ3_KT-00001-of-00002.gguf
6fd25e2388ae16837c1634ca3b06486fce8b7db11b216a970f6c008d56d83e33  Qwen3.8-Flash-Next-ik_llama-IQ3_KT-00002-of-00002.gguf
c153dde533ca2e578688bc95677e0273737adf2de2df5919d40cbfe080528fbb  Qwen3.8-Flash-Next-ik_llama-IQ4_KT-00001-of-00002.gguf
398978ac7ca29cd9c5328057468846b61a2375b69ab9d96d880927ceb3bb441c  Qwen3.8-Flash-Next-ik_llama-IQ4_KT-00002-of-00002.gguf

Downloads last month: 6,236

GGUF

Model size

177B params

Architecture

qwen4exp

Hardware compatibility

Model tree for ji-farthing/Qwen3.8-Flash-Next-ik-llama-GGUF

Base model

Qwen/Qwen3.8-Flash-Next

Quantized

(210)

this model