Instructions to use Ninnix96/Qwengram-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Ninnix96/Qwengram-4B 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 Ninnix96/Qwengram-4B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Ninnix96/Qwengram-4B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Ninnix96/Qwengram-4B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Ninnix96/Qwengram-4B: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 Ninnix96/Qwengram-4B:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Ninnix96/Qwengram-4B: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 Ninnix96/Qwengram-4B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Ninnix96/Qwengram-4B:Q4_K_M

Use Docker

docker model run hf.co/Ninnix96/Qwengram-4B:Q4_K_M

LM Studio
Jan

vLLM

How to use Ninnix96/Qwengram-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ninnix96/Qwengram-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ninnix96/Qwengram-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Ninnix96/Qwengram-4B:Q4_K_M

Ollama
How to use Ninnix96/Qwengram-4B with Ollama:
```
ollama run hf.co/Ninnix96/Qwengram-4B:Q4_K_M
```
Unsloth Desktop

How to use Ninnix96/Qwengram-4B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Ninnix96/Qwengram-4B: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": "Ninnix96/Qwengram-4B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Ninnix96/Qwengram-4B with Docker Model Runner:
```
docker model run hf.co/Ninnix96/Qwengram-4B:Q4_K_M
```

Lemonade

How to use Ninnix96/Qwengram-4B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Ninnix96/Qwengram-4B:Q4_K_M

Run and chat with the model

lemonade run user.Qwengram-4B-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Ninnix96/Qwengram-4B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Ninnix96/Qwengram-4B: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 Ninnix96/Qwengram-4B:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Ninnix96/Qwengram-4B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Ninnix96/Qwengram-4B: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 "Ninnix96/Qwengram-4B: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"

Qwengram-4B / runtime /README.md

Ninnix96

Release Qwengram 4B REAL-15M plus linear750

feec8ea verified 7 days ago

preview code

Raw History Blame Contribute Delete

3.4 kB

Qwengram-4B GGUF runtime validation

Canonical checkpoint: REAL-15M + linear750. See the frozen decision for the independent 10M/15M comparison.

Precision	Stock NLL	Qwengram NLL	Reader gain [95% CI]	Gain retention [95% CI]	Perplexity reduction vs stock
BF16	2.225750	2.198449	0.027301 [0.022650, 0.031917]	100%	2.69%
Q8_0	2.226990	2.199469	0.027521 [0.022892, 0.032133]	100.8% [97.1%, 104.7%]	2.71%
Q6_K	2.230865	2.200943	0.029922 [0.023478, 0.036439]	109.6% [99.4%, 118.6%]	2.95%
Q4_K_M	2.262792	2.231828	0.030964 [0.025884, 0.036537]	113.4% [100.6%, 127.6%]	3.05%

Matched test

The matched CPU test scores 8,128 tokens from the first 64 consecutive 256-token WikiText-2 raw test chunks, scoring the last 127 tokens per chunk. All runs use eight threads and context/batch/microbatch 256, with no warmup. Reader gain is NLL(stock) - NLL(Qwengram); retention divides each quantized gain by the BF16 gain. Paired 95% intervals use 10,000 resamples of 16 consecutive four-chunk blocks, seed 1234. The external PLE is Ivan Fioravanti's Q4_1 sidecar.

BF16 has the lowest absolute Qwengram NLL in this test. Gain retention measures the added PLE benefit within each precision.

All eight stock/Qwengram runs were measured with the same patched runtime. Per-chunk scores and model hashes are in results.json; commands, timings and logs are in logs.

All 441 backbone tensors and nine tokenizer fields match exactly within each stock/Qwengram pair. All 11 reader/arbiter tensors remain bit-exact FP32 across BF16, Q8_0, Q6_K and Q4_K_M and match the verified reader and matching arbiter. See packaging and verification.

The PLE sidecar check compares 4,096 addressed rows with independent reference addressing and dequantization. Full prefill, split prefill, token decode, EOS boundaries and repeated reset have max absolute difference zero. The PLE remains unchanged.

This CPU test uses a quantized Q4_1 PLE and a WikiText-2 slice. The frozen Kaggle study uses the original FP8 PLE and different evaluation streams. These scores are separate benchmarks.

Generation checks

On the tested AMD BC-250, BF16, Q8_0, Q4_K_M with full Vulkan offload (-ngl 99) matched the corresponding CPU eight-token greedy continuation for The capital of France is. These short checks do not establish broad GPU parity. Q6_K differed at full offload but matched with -ngl 33, which keeps the first decoder layer on CPU. Use -ngl 33 for Q6_K on this tested device, or CPU (-ngl 0). The fallback is also a short generation check.

The loader rejected missing and invalid PLE sidecars and conflicting injection indices. The prior 0.8B and 2B Q8_0 CPU continuations remain exact with the updated runtime. See smoke.json, Q6 output, Q6 fallback and legacy checks.

Runtime source

Published fork: 3616a858f2326e87ad8b48e1341a4e341b3dad73. The tested base commit, source patch and binary hashes are recorded in verification.json; the source patch reproduces the tested source.