qwen3-4b-lowram-bench

ANTIMEME SYSTEMWORKS // CREATIVE UTILITIES DIVISION

Qwen3-4B / Low-RAM benchmark

8 quantizations. One CPU setup. Quality, file size and throughput in the same record.

AM-RS-2026 // QWEN3-4B // V1.0_BUILD

---

GGUF conversions and quantizations of Qwen/Qwen3-4B, with measured WikiText-2 perplexity and CPU throughput. The files contain no fine-tuning. Compare the recorded setup and uncertainty before transferring these results to another machine.

01 / 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:

1. Machine-readable results (data/results.tsv)
2. Exact methodology; pinned llama.cpp build + commit, fixed corpus, fixed context, fixed threads, fixed seed
3. A CPU benchmark you can re-run (bench/run_bench.sh)
4. 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.

02 / 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).

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.

03 / 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):

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 reduces the perplexity gap (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.

04 / Benchmarks

Eight quantizations, CPU-only, 16 threads. The perplexity axis is truncated to show the measured differences; throughput starts at zero.

Open the interactive quant picker. Select a row for the exact GGUF download. File size is measured; the picker's RAM figure is a rule-of-thumb estimate.

05 / Files

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

06 / 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

Use a compatible llama.cpp-based client, such as llama-cli or llama-server, with support for this model and quantization.

07 / 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.

08 / 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).

09 / 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

10 / 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)

11 / 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.

---

ANTIMEME SYSTEMWORKS // CREATIVE UTILITIES DIVISION // P.01