Instructions to use b4ph/qwen3-4b-lowram-bench with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use b4ph/qwen3-4b-lowram-bench 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 b4ph/qwen3-4b-lowram-bench:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf b4ph/qwen3-4b-lowram-bench:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf b4ph/qwen3-4b-lowram-bench:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf b4ph/qwen3-4b-lowram-bench: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 b4ph/qwen3-4b-lowram-bench:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf b4ph/qwen3-4b-lowram-bench: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 b4ph/qwen3-4b-lowram-bench:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf b4ph/qwen3-4b-lowram-bench:Q4_K_M

Use Docker

docker model run hf.co/b4ph/qwen3-4b-lowram-bench:Q4_K_M

LM Studio
Jan

vLLM

How to use b4ph/qwen3-4b-lowram-bench with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "b4ph/qwen3-4b-lowram-bench"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "b4ph/qwen3-4b-lowram-bench",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/b4ph/qwen3-4b-lowram-bench:Q4_K_M

Ollama
How to use b4ph/qwen3-4b-lowram-bench with Ollama:
```
ollama run hf.co/b4ph/qwen3-4b-lowram-bench:Q4_K_M
```
Unsloth Desktop

How to use b4ph/qwen3-4b-lowram-bench with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf b4ph/qwen3-4b-lowram-bench: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": "b4ph/qwen3-4b-lowram-bench:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use b4ph/qwen3-4b-lowram-bench with Docker Model Runner:
```
docker model run hf.co/b4ph/qwen3-4b-lowram-bench:Q4_K_M
```

Lemonade

How to use b4ph/qwen3-4b-lowram-bench with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull b4ph/qwen3-4b-lowram-bench:Q4_K_M

Run and chat with the model

lemonade run user.qwen3-4b-lowram-bench-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use b4ph/qwen3-4b-lowram-bench with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf b4ph/qwen3-4b-lowram-bench: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 b4ph/qwen3-4b-lowram-bench:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use b4ph/qwen3-4b-lowram-bench with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf b4ph/qwen3-4b-lowram-bench: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 "b4ph/qwen3-4b-lowram-bench: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"

Qwen3-4B — Low-RAM Quantization Benchmark (reproducible, CPU-only)

A reproducible, transparent benchmark of the base Qwen/Qwen3-4B model across the standard llama.cpp K-quantization ladder, plus an imatrix comparison — measured on a single real CPU, with machine-readable data and exact reproduction steps.

No training, no fine-tuning, no modifications. These are pure GGUF conversions and quantizations of the Apache-2.0 base model. The value of this repo is the benchmark methodology and the numbers — not the weights.

Why this repo exists

Most GGUF model cards on the Hub give subjective quality labels ("very high quality, recommended") with no reproducible numbers, and the few that publish perplexity are single-setup with no way to reproduce them. This repo fills that gap:

Machine-readable results (data/results.tsv)
Exact methodology — pinned llama.cpp build + commit, fixed corpus, fixed context, fixed threads, fixed seed
A CPU benchmark you can re-run (bench/run_bench.sh)
The imatrix question answered for base Qwen3-4B — the one comparison that separates "good quant" from "best quant," independently reproduced here on a real CPU.

Results

Perplexity on the canonical WikiText-2 test set (the exact corpus llama.cpp's own CI uses — so these numbers are directly comparable to the ecosystem's), context 2048, 16 threads, all CPU (-ngl 0), seed 1. Throughput via llama-bench (512-token prompt processing pp512 and 128-token generation tg128, 5 repeats, mean reported).

Quant	File size	Perplexity (WikiText-2)	Δ vs F16	pp512 (t/s)	tg128 (t/s)
F16 (BF16)	8.05 GB	13.4304	—	357.6	2.52
Q8_0	4.28 GB	13.4409	+0.0105	1355.9	5.04
Q6_K	3.31 GB	13.4404	+0.0099	1523.2	5.93
Q5_K_M	2.89 GB	13.5362	+0.1058	1917.1	5.64
Q4_K_M	2.50 GB	13.7304	+0.3000	2064.8	6.50
Q4_K_M + imatrix	2.50 GB	13.6760	+0.2456	1965.3	6.41
Q3_K_M	2.08 GB	15.6641	+2.2337	2142.5	6.56
Q3_K_M + imatrix	2.08 GB	14.8237	+1.3933	2308.2	5.89

Reading the table: Q8_0/Q6_K are essentially lossless (+0.01 PPL). Q4_K_M is the knee of the curve — 2.5 GB (31% of BF16) for +0.30 PPL, and it's the fastest generator of the non-Q3 set. Q3_K_M is a real quality cliff (+2.23 PPL), which is exactly where imatrix earns its keep.

The imatrix question

imatrix ("importance matrix") quantization uses a calibration set (wikitext-2 wiki.train.raw, 11 MB — standard practice) to learn which weight directions matter most, so the quantizer spends its bits where they count. Measured here for base Qwen3-4B (the delta vs the same non-imatrix quant):

Quant	PPL without imatrix	PPL with imatrix	Δ (imatrix gain)
Q4_K_M	13.7304	13.6760	−0.054 (small)
Q3_K_M	15.6641	14.8237	−0.840 (large)

The pattern is the well-documented one, now reproduced at this scale: imatrix pays off disproportionately at low bit-counts, where every bit matters. At Q4_K_M the gain is real but small; at Q3_K_M — where the plain quant is already 2.23 PPL worse than F16 — imatrix recovers most of the cliff (back to +1.39). The cost is a one-time calibration pass (~3 h on this CPU; trivial on GPU) and a ~2.5 MB .dat file. Rule of thumb: run imatrix when you drop to Q3 or below. The two imatrix files let you A/B this yourself.

Benchmarks

Size vs accuracy and size vs speed (CPU-only, 16 threads). Light/dark variants render automatically from your Hub theme.

Qwen3-4B size vs accuracy and speed (light)

Qwen3-4B size vs accuracy and speed (dark)

Files

File	Quant	Size	Notes
`Qwen3-4B-F16.gguf`	F16	8.05 GB	full-precision baseline
`Qwen3-4B-Q8_0.gguf`	Q8_0	4.28 GB	near-lossless
`Qwen3-4B-Q6_K.gguf`	Q6_K	3.31 GB	K-quant (super-block scaling)
`Qwen3-4B-Q5_K_M.gguf`	Q5_K_M	2.89 GB	K-quant, "M" = mixed precision
`Qwen3-4B-Q4_K_M.gguf`	Q4_K_M	2.50 GB	workhorse — runs in ~3 GB RAM
`Qwen3-4B-Q4_K_M_imatrix.gguf`	Q4_K_M+im	2.50 GB	imatrix-quantized
`Qwen3-4B-Q3_K_M.gguf`	Q3_K_M	2.08 GB	smallest plain quant
`Qwen3-4B-Q3_K_M_imatrix.gguf`	Q3_K_M+im	2.08 GB	imatrix-quantized

Machine-readable results: data/results.tsv · Eval corpus: data/wiki.test.raw · Harness: bench/run_bench.sh

How to use

pip install -U "huggingface_hub[cli]"
# grab just the file you want
hf download b4ph/qwen3-4b-lowram-bench --include "Qwen3-4B-Q4_K_M.gguf" --local-dir ./models

Run locally (llama.cpp):

llama-cli -m models/Qwen3-4B-Q4_K_M.gguf -c 4096

Any llama.cpp-based client works: LM Studio, Ollama (ollama import), llama-server, text-generation-webui, etc.

How to reproduce the numbers

Harness: llama.cpp build 10516, commit b95502ba9 (the exact build used — the quantizer's behavior is version-sensitive, so pinning matters).

Hardware: 16-core x86-64 CPU (AVX-512 + VNNI), 61.6 GB RAM. GPU present (RTX 5090) but unused (-ngl 0) — all numbers are pure-CPU, so they're reproducible on any comparable CPU box.

# quality axis — fixed corpus, fixed ctx, fixed seed
llama-perplexity -m Qwen3-4B-<QUANT>.gguf \
  -f wiki.test.raw -c 2048 -t 16 -ngl 0 -s 1

# speed axis — 512-token prompt + 128-token generation, 5 repeats
llama-bench -m Qwen3-4B-<QUANT>.gguf -p 512 -n 128 -t 16 -ngl 0 -r 5

To add imatrix quants (as done for Q4_K_M and Q3_K_M here):

# 1. generate the importance matrix from a calibration set
llama-imatrix -m Qwen3-4B-F16.gguf -f wiki.train.raw -c 2048 -t 16 -ngl 0 -o imatrix.dat
# 2. quantize using it (note: long --imatrix flag)
llama-quantize --imatrix imatrix.dat Qwen3-4B-F16.gguf Qwen3-4B-Q4_K_M_imatrix.gguf Q4_K_M

The full loop (perplexity + speed for every quant, TSV output) is in bench/run_bench.sh.

Base model

Converted from Qwen/Qwen3-4B via llama.cpp's convert_hf_to_gguf.py (at the matching build commit). Model: 36 layers, hidden 2560, GQA 32/8 heads, vocab 151936, 4.0B params, natively 262,144 context (RoPE-scaled).

Appropriate uses

CPU-only / low-RAM deployment of Qwen3-4B (laptops, SBCs, servers without GPUs)
Quantization selection: pick the smallest quant whose perplexity delta you accept
Benchmarking/teaching material for the size↔quality tradeoff

Out-of-scope uses

Any claim that these quants were trained, fine-tuned, or otherwise modified
Using these numbers to compare against other models (different corpus settings would be needed for a fair cross-model eval)
High-stakes / safety-critical deployment — Qwen3-4B's own limitations apply (see the base model card)

Licensing

Base model: Apache 2.0 (per Qwen/Qwen3-4B's LICENSE, which is carried in this repo as LICENSE)
These GGUFs are derivative works of the Apache-2.0 base → Apache 2.0
Eval corpus: WikiText-2 (research benchmark corpus)

Community benchmark, independent of Qwen/Alibaba. Not affiliated with or endorsed by Alibaba Cloud.

Downloads last month: 86

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Model tree for b4ph/qwen3-4b-lowram-bench

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Quantized

(313)

this model

b4ph
/

qwen3-4b-lowram-bench

Qwen3-4B — Low-RAM Quantization Benchmark (reproducible, CPU-only)

Why this repo exists

Results

The imatrix question

Benchmarks

Files

How to use

How to reproduce the numbers

Base model

Appropriate uses

Out-of-scope uses

Licensing

Model tree for b4ph/qwen3-4b-lowram-bench

Space using b4ph/qwen3-4b-lowram-bench 1