---

license: apache-2.0

base_model: Qwen/Qwen3.5-0.8B

base_model_relation: quantized

pipeline_tag: text-generation

tags:

- gguf

- quantized

- qwen3.5

- qat

- llama.cpp

---

Qwen3.5-0.8B (Yooz QAT, GGUF Q4_0)

> Context limit: 8K. This build is validated healthy up to 8,192 tokens.

> Beyond that (tested to 32K on the MLX build of the same weights) it

> regresses on multi-hop retrieval -- NIAH single-needle drops to 0.03

> accuracy, below the untrained 4-bit baseline. **Do not use this build

> past 8K context.** If you need long context at this size class, use

> Qwen3.5-4B-qat-GGUF

> instead, which is clean through 32K. Full breakdown in Validation below.

What this is

The GGUF (Q4_0) release of Yooz Labs' quantization-aware-trained (QAT)

Qwen3.5-0.8B -- the same weights as

Qwen3.5-0.8B-qat-lean-4bit-mlx,

converted for llama.cpp. Stock 4-bit PTQ builds of this model damage its

aggregation abilities (variable-tracking collapses from a healthy score to

near-zero 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 instead of trading one task for another.

Provenance

• Base checkpoint: Qwen/Qwen3.5-0.8B
• 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). Training recipes are not

published.

• Conversion: llama.cpp convert_hf_to_gguf.py + llama-quantize

to Q4_0 (build e3546c7). The converter's block count is set to 24

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

Qwen3.5-0.8B-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 | 1.00 |

| Variable tracking | 8K | 0.23 | 0.78 |

| Frequent words | 8K | 0.28 | 0.36 |

| MRCR 2-needle | 8K | 0.80 | 0.67 |

| NIAH single | 32K | 1.00 | 0.03 |

| NIAH multikey | 32K | 1.00 | 0.50 |

| Variable tracking | 32K | 0.49 | 0.16 |

| Frequent words | 32K | 0.33 | 0.31 |

| MRCR 2-needle | 32K | 0.67 | 0.64 |

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

RTX 4090 (13th Gen Intel Core i9-13900K, Ubuntu 24.04, llama.cpp build

e3546c7):

| Task | Context | Yooz QAT 4-bit (this GGUF file) |

|---|---|---:|

| Variable tracking | 8K | 0.88 |

| MRCR 2-needle | 8K | 0.69 |

Consistent with the MLX numbers above -- the fix survives re-quantization

to a different format. As with the MLX build, do not use this file past

8K: the 32K regression was characterized on the MLX build of the same

weights and there is no reason to expect Q4_0 fixes it.

Usage

llama-cli -m Qwen3.5-0.8B-qat-Q4_0.gguf -p "List the first five prime numbers." -n 64

Or with llama-server for an OpenAI-compatible endpoint:

llama-server -m Qwen3.5-0.8B-qat-Q4_0.gguf --port 8080

An MLX build of the same QAT weights is available at

YoozLabs/Qwen3.5-0.8B-qat-lean-4bit-mlx.

About Yooz Labs

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.

Questions, issues, or corrections: dev@yooz.info

License and attribution

Apache 2.0, inherited from the base checkpoint Qwen/Qwen3.5-0.8B. Qwen is

a model family by Alibaba Cloud; this is an independent, unofficial QAT

build, not distributed, endorsed, or certified by Alibaba.