The GGUF (Q4_0) release of Yooz Labs' quantization-aware-trained (QAT)
Qwen3.5-4B -- the same weights as
Qwen3.5-4B-qat-lean-4bit-mlx,
converted for llama.cpp. Stock 4-bit PTQ builds of this model break its
aggregation abilities outright (variable-tracking collapses to 0.00 on
the MLX product grid); this build is trained to tolerate 4-bit
quantization via knowledge distillation from the full-precision bf16
model, so the quantized weights reproduce the original model's behavior
-- aggregation AND multi-turn recall -- at both short and long context.
Provenance
Base checkpoint:Qwen/Qwen3.5-4B
QAT training: Yooz Labs, 2026-07 -- knowledge distillation from the
bf16 base model under fake-quantization matching the MLX 4-bit
group-64 grid (the training target; this GGUF is the same trained
weights re-quantized to Q4_0 for llama.cpp). Full-parameter training at
this size. Training recipes are not published.
Conversion:llama.cppconvert_hf_to_gguf.py + llama-quantize
to Q4_0 (build e3546c7). The converter's block count is set to 32
layers, matching this model's actual depth.
Harness:benchmarks/ (infinite repo, Yooz Labs).
Validation
The comprehensive A/B (needle-in-a-haystack, RULER-style aggregation,
MRCR) was run on the MLX build of these exact weights, since that is
where the full n=30, two-context-length harness lives; see
Qwen3.5-4B-qat-lean-4bit-mlx
for the full table and chart. Summary:
Task
Context
4-bit PTQ, same grid
Yooz QAT 4-bit
NIAH single
8K
1.00
1.00
NIAH multikey
8K
1.00
0.97
Variable tracking
8K
0.00
0.97
Frequent words
8K
0.12
0.83
MRCR 2-needle
8K
0.97
0.97
NIAH single
32K
0.97
1.00
NIAH multikey
32K
1.00
0.97
Variable tracking
32K
0.89
0.99
Frequent words
32K
0.22
0.77
MRCR 2-needle
32K
0.96
0.96
This GGUF file was separately spot-checked through the actual
llama.cpp Q4_0 path (not just the MLX grid) at 8K context, n=30, on an
A100-SXM4-80GB (AMD EPYC 7713, Ubuntu 22.04, llama.cpp build e3546c7):
Task
Context
Yooz QAT 4-bit (this GGUF file)
Variable tracking
8K
0.96
Frequent words
8K
0.62
MRCR 2-needle
8K
0.94
Consistent with the MLX numbers above -- the fix survives re-quantization
to a different format.
Usage
llama-cli -m Qwen3.5-4B-qat-Q4_0.gguf -p "List the first five prime numbers." -n 64
Or with llama-server for an OpenAI-compatible endpoint:
Maintained by Yooz Labs -- sovereign, on-device AI for
macOS. Everything we ship runs locally; nothing goes to a cloud. This build
comes out of our long-context research line, where we found and quantified
the PTQ aggregation crater -- and then trained it away.
Apache 2.0, inherited from the base checkpoint Qwen/Qwen3.5-4B. Qwen is
a model family by Alibaba Cloud; this is an independent, unofficial QAT
build, not distributed, endorsed, or certified by Alibaba.
