Instructions to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF",
	filename="Seed-OSS-36B-Instruct-unsloth-MXFP4_MOE-output_q6_K-router_gate_emb_q6_K.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K
# Run inference directly in the terminal:
llama-cli -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K
# Run inference directly in the terminal:
llama-cli -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

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 magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K
# Run inference directly in the terminal:
./llama-cli -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

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 magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K
# Run inference directly in the terminal:
./build/bin/llama-cli -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

Use Docker

docker model run hf.co/magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

LM Studio
Jan

vLLM

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-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": "magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

Ollama
How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with Ollama:
```
ollama run hf.co/magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K
```

Unsloth Studio new

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF to start chatting

Pi new

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

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 magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

Run Hermes

hermes

Docker Model Runner
How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with Docker Model Runner:
```
docker model run hf.co/magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K
```

Lemonade

How to use magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull magiccodingman/Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF:Q6_K

Run and chat with the model

lemonade run user.Seed-OSS-36B-Instruct-Unsloth-MXFP4-Hybrid-GGUF-Q6_K

List all available models

lemonade list

DEPRECIATED!

This model was surpassed by the new MagicQuant hybrids. The collection can be found here:

https://huggingface.co/collections/magiccodingman/magic-quant

Use the new version. This shown MXFP4 hybrid is no longer viable in comparison, nor really useable. The data collected is good for understanding and research, but it's not as good for real use.

Unsloth - Seed OSS 36B Instruct MXFP4 Hybrid GGUF

Dense model utilizing MXFP4_MOE with hybrid weights on a dense model. Achieving interesting results that show smaller file size, more TPS, and near lossless precision.

Use The Following Models!

Stats compared against the standard Q8_0 (precision loss still compared to F16)

MXFP4_MOE-output_q6_K-router_gate_emb_q6_K

8.8% smaller than Q8 • 21.82 TPS • 0.053% precision loss

Surprising Result

MXFP4_MOE

11.7% smaller than Q4_K_M • 20.94 TPS • 2.694% precision loss

The MXFP4_MOE model got a lower file size than the Q4_K_M and slightly better precision (though at the cost of less TPS than the Q4_K_M). This result has been an outlier compared to any other test of mine, but it's why this model was included.

This repository contains a set of hybrid MXFP4 quantized GGUF models designed to explore a surprising discovery:

A carefully targeted combination of MXFP4 + high-precision embeddings/output weights can deliver near-Q8 accuracy with Q4–Q6 level throughput and smaller file sizes than Q8.

Unlike pure MXFP4, which heavily degrades dense models. This hybrid method selectively protects tensors that matter most for semantic stability, while allowing MXFP4 to accelerate everything else.

This is experimental. And should be treated as such. I am more than encouraging people to use this model and leave feedback! Though precision loss seemed near lossless, did the hybrid models act strange in certain situations? Worse or better on some topics compared to the original model? Did it do better/worse overall on everything? I'd love to hear back from others!

The Magic Model

This model achieved:

File size reduction compared to the Q8_0

Better precision loss scores than the pure Q6_K

Achieving noticeably better TPS than a Q4_K_M

I have personally deemed this in the category of "Q7.5" quantization.

MXFP4_MOE-output_q6_K-router_gate_emb_q6_K

(8.8% smaller than Q8 • 21.82 TPS • 0.053% precision loss )

This version created beat out everything in every way in the MXFP4 hybrid family created. Out of the batch, this MXFP4 hybrid was the only worth considering to utilize.

The following was the conversion script:

llama-quantize \
  --tensor-type token_embd.weight=Q6_K \
  --tensor-type output.weight=Q6_K \
  --tensor-type 'router.*'=Q6_K \
  --tensor-type 'gate.*'=Q6_K \
  "Path_To_F16_GGUF.gguf" \
  "Path_To_GGUF.gguf" \
  mxfp4_moe

MXFP4_MOE Hybrid Naming Scheme & Synopsis

Multiple different combinations of converted models were created. The results were interesting to say the least. The following table will explain my naming scheme to what was done to the model to create it.

Suffix Example	Meaning
`MXFP4_MOE`	Pure MXFP4 pipeline
`MXFP4_MOE-Q8`	Embedding/output in Q8_0
`MXFP4_MOE-F16`	Embedding/output in F16
`output_mxfp4-embd_q8`	Output → MXFP4, Embedding → Q8
`output_mxfp4-router_gate_emb_q5_K`	Output → MXFP4, Emb/Router/Gate → Q5_K
`MXFP4_MOE-Q6_K`	Both embedding + output in Q6_K
`Q8_0`, `Q6_K`, `Q4_K_M`	Pure model-wide quantizations

The results achieved were interesting to say the least. It was a brute force game of mass creating models with hybrid methods to find combinations that didn't cause too much noise and paired well with MXFP4.

This repo showcases the converted models, whether good or bad that was created. But, I have been testing other models in different combinations as well. The winning hybrid combinations shown in this repo DOES NOT always equate to the same results on different models.

Some models do better or worse with different kinds of combinations. It depends if it's dense, MOE, and much more. Many times the results surprise me. Many models no matter the combination will not play nice with MXFP4. At least with the methods shown here.

Benchmark Methodology

All models were tested with a unified automated harness using llama.cpp tools.

Included tests:

Throughput:
llama-bench with descending GPU offload (-ngl 35 → 0) and automatic OOM retry.
Highest successful TPS is recorded.
Perplexity:
Three domains: general, code, math.
Each uses an auto-generated corpus of ~32k tokens.
Perplexity is computed with llama-perplexity at 2048-token context.
Same GPU retry logic as above.
Precision loss:
Each model is compared to its family F16 baseline.
Precision-loss % is computed for all PPL domains, plus an averaged score.
Models are ranked by this metric.

Table - Overview of Results

Comparing to F16.

model_name	size_reduction	tps_change
MXFP4_MOE-Q8	46.87%	61.73%
Q8_0	46.87%	66.72%
MXFP4_MOE-F16	40.46%	41.91%
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	51.57%	84.76%
MXFP4_MOE-Q6_K	48.52%	66.55%
Q6_K	58.98%	97.63%
Q5_K_M	64.6%	90.69%
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	53.21%	82.98%
MXFP4_MOE-output_mxfp4-embd_q6_K	50.19%	74.01%
MXFP4_MOE-output_mxfp4-embd_q8	49.92%	64.94%
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	49.92%	74.09%
MXFP4_MOE	73.42%	77.31%
Q4_K_M	69.9%	132.18%

All percentages compared against the selected family F16 baseline.

Table - File Size + TPS + Avg Precision Loss

model_name	file_size_gb	bench_tps	avg_prec_loss
F16	67.35	11.81	0
MXFP4_MOE-Q8	35.78	19.1	0.0171
Q8_0	35.78	19.69	0.0171
MXFP4_MOE-F16	40.1	16.76	0.0215
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	32.62	21.82	0.053
MXFP4_MOE-Q6_K	34.67	19.67	0.0566
Q6_K	27.63	23.34	0.1651
Q5_K_M	23.84	22.52	0.2512
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	31.51	21.61	1.0377
MXFP4_MOE-output_mxfp4-embd_q6_K	33.55	20.55	1.0464
MXFP4_MOE-output_mxfp4-embd_q8	33.73	19.48	1.0473
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	33.73	20.56	1.0473
MXFP4_MOE	17.9	20.94	2.694
Q4_K_M	20.27	27.42	2.8138

Bench NGL was 35
Utilized CUDA

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
F16	6.8905	0.1681	1.4129	0.0095	5.4475	0.121
MXFP4_MOE-Q8	6.8866	0.1679	1.413	0.0095	5.4474	0.121
Q8_0	6.8866	0.1679	1.413	0.0095	5.4474	0.121
MXFP4_MOE-F16	6.8893	0.1679	1.4132	0.0095	5.4508	0.1211
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	6.8932	0.1682	1.4127	0.0095	5.4548	0.1213
MXFP4_MOE-Q6_K	6.8946	0.1682	1.4128	0.0095	5.4539	0.1213
Q6_K	6.9012	0.1685	1.4135	0.0095	5.4637	0.1218
Q5_K_M	6.9071	0.1685	1.4168	0.0096	5.4604	0.1212
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	6.9647	0.169	1.4196	0.0095	5.5326	0.1227
MXFP4_MOE-output_mxfp4-embd_q6_K	6.9649	0.1691	1.4199	0.0095	5.5327	0.1226
MXFP4_MOE-output_mxfp4-embd_q8	6.9638	0.1691	1.4198	0.0095	5.5341	0.1227
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	6.9638	0.1691	1.4198	0.0095	5.5341	0.1227
MXFP4_MOE	7.1007	0.1728	1.4351	0.0097	5.636	0.1239
Q4_K_M	7.0964	0.1759	1.4235	0.0098	5.7037	0.1303

gen = ppl_general
gen_er = ppl_general_error
code = ppl_code
code_er = ppl_code_error
math = ppl_math
math_er = ppl_math_error

Table - Precision Loss Columns

model_name	loss_general	loss_code	loss_math
F16	0	0	0
MXFP4_MOE-Q8	-0.0566	0.0071	-0.0018
Q8_0	-0.0566	0.0071	-0.0018
MXFP4_MOE-F16	-0.0174	0.0212	0.0606
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	0.0392	-0.0142	0.134
MXFP4_MOE-Q6_K	0.0595	-0.0071	0.1175
Q6_K	0.1553	0.0425	0.2974
Q5_K_M	0.2409	0.276	0.2368
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	1.0768	0.4742	1.5622
MXFP4_MOE-output_mxfp4-embd_q6_K	1.0797	0.4954	1.564
MXFP4_MOE-output_mxfp4-embd_q8	1.0638	0.4884	1.5897
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	1.0638	0.4884	1.5897
MXFP4_MOE	3.0506	1.5712	3.4603
Q4_K_M	2.9882	0.7502	4.7031