Mixed-precision GGUF quantizations of Qwen3.5-4B — the first BitClass release of this model. BitClass3 keeps the Hessian-sensitivity front-end to set each level's bit budget, but hands the per-tensor allocation to an error-minimizing solver (built on llama.cpp's --target-bpw) that distributes bits across every tensor — including the hybrid DeltaNet/SSM tensors — to minimize imatrix-weighted quantization error at the target size.

Available Quantizations

KLD vs the BF16 source is the primary quality metric (mean and the robust 99.9th percentile); wikitext-2 perplexity is reported alongside.

File	BPW	Size	wiki PPL ↓	KL-mean ↓	KL-99.9% ↓	Use Case
`Qwen3.5-4B-Q8_0.gguf`	8.5	4.48 GB	8.638	0.0028	0.073	Near-lossless reference
`Qwen3.5-4B-Q6_K.gguf`	6.2	3.25 GB	8.767	0.0125	0.396	High quality
`Qwen3.5-4B-Q5_K_M.gguf`	5.1	2.71 GB	9.176	0.0255	0.803	Balanced quality and size
`Qwen3.5-4B-Q4_K_M.gguf`	5.0	2.61 GB	9.246	0.0292	0.952	Best quality-to-size ratio
`Qwen3.5-4B-Q3_K_S.gguf`	3.8	1.98 GB	9.358	0.0978	3.583	Maximum compression

Recommended: Q4_K_M — KL-mean 0.029 at 2.61 GB.

How It Compares

Same harness, same metrics, against a fixed per-suffix LP-recipe baseline at matched BPW (within 0.5%) — the head-to-head that motivated BitClass3's allocator:

Level	LP recipe KL-mean	BitClass3 KL-mean
Q3_K_S	0.1525	0.0978
Q4_K_M	0.0352	0.0292
Q5_K_M	0.0258	0.0255

The allocator wins KL divergence at every level — largest at aggressive quantization (Q3_K_S: −36% KL-mean) — and our internal worst-token measurements improve at every level as well. A key reason: it allocates bits to the hybrid DeltaNet tensors (attn_qkv, attn_gate, ssm_*) that a standard 7-suffix recipe leaves at the base type — and our tensor-health scan shows those exact tensors are the statistical outliers of this architecture.

Key Sensitivity Findings (Qwen3.5-4B)

blk.3 (early layer) is most sensitive — the same early-layer pattern as Qwen3.5-9B, and the opposite of dense Qwen3-4B-Instruct (blk.34). The hybrid Qwen3.5 family concentrates sensitivity early.
Attention K projections are consistently ≥ V in sensitivity.
DeltaNet/SSM tensors are distribution outliers (high kurtosis ssm_conv1d, shifted ssm_alpha/beta/out, attn_qkv, attn_gate vs same-role peers) — covering them in the allocation matters; ssm_conv1d itself is kept at F32 by llama.cpp.

How It Works

Hessian sensitivity — compute H_diag = mean(X²) per layer on calibration data; this sets each level's overall bit budget.
Error-minimizing per-tensor allocation — an imatrix-weighted solver (llama.cpp --target-bpw) assigns a quant type to every tensor to minimize total quantization error at the target BPW, covering attention, FFN, and the hybrid DeltaNet/SSM tensors.
imatrix — importance matrix computed over wikitext guides the per-tensor error.
GGUF export — produced with stock llama-quantize.

Usage

bash

hf download sh111111111111111/Qwen3.5-4B-BitClass3-GGUF \
    Qwen3.5-4B-Q4_K_M.gguf --local-dir .

llama-cli    -m Qwen3.5-4B-Q4_K_M.gguf -cnv
llama-server -m Qwen3.5-4B-Q4_K_M.gguf --port 8080

Note: Qwen3.5 GGUFs are not currently runnable in Ollama (vision/mmproj handling is not yet supported there); use llama.cpp or LM Studio.

Benchmark Details

NVIDIA GB10 ATOM (128 GB unified memory, aarch64). llama.cpp with --target-bpw (PR #15550). KLD via llama-perplexity --kl-divergence against BF16-source logits over wikitext-2 (mean / median / 99.9th percentile reported; the single-token KL-max is omitted as an unstable order statistic). wikitext-2 PPL via llama-perplexity -c 2048. Downstream (HellaSwag / WinoGrande / ARC / MMLU) tracked internally.

Disclaimer

Independent project. Not affiliated with or endorsed by Qwen, Unsloth, ByteShape, Bartowski, or llama.cpp. Competitor figures are from our own benchmark harness and may differ from those projects' self-reported numbers; competitor file sizes reflect the revision we tested and may since have changed.

License

Apache 2.0, inherited from Qwen3.5-4B.