Qwen3.8-27B quantization: Q4, Q5 or W4A16?

On the home RTX 3090, quantization selected a complete operating profile. Q5 consumed more space and constrained long context. Q4 left KV cache more room. W4A16 AutoRound produced the strongest whole-system result in one task on another engine, but that run cannot isolate weight quality.

When a supplier promises “almost no quality loss”, ask for the result on your tasks and the exact artifact hash.

What gets compressed

A 27B model in BF16 needs roughly 55 GB for weights. Quantization stores some values with fewer bits, reducing memory and memory bandwidth while introducing approximation.

In practice, a profile separately chooses:

• model-weight precision;
• KV-cache precision for prior context;
• precision of the additional MTP head;
• a format and kernels supported by the server.

“4-bit” is therefore not one technology. GGUF Q4_K_M, AutoRound W4A16 and NVFP4 use different procedures, precision layouts and runtimes.

Official Unsloth GGUF sizes

The Qwen3.8-27B GGUF repository lists:

Our 60k llama.cpp profile used Q5_K_M and Q8 cache. The 120k and 250k profiles used Q4_K_M. At 250k, cache dropped to Q4 and batch settings became much smaller. The longer window was purchased with several compromises.

Four profiles that actually fit in 24 GB

The first three rows came from the same recall test. The final row came from a different long agent task, so it is not a direct memory benchmark against GGUF. It is the observed peak for the profile selected for further work.

The 250k profile fit with less than 1 GiB free, a smaller batch and Q4 cache. It was an experimental maximum-window profile. The measurement does not show comfortable capacity for several sessions. The server still had one slot.

How W4A16 differs

W4A16 roughly means 4-bit weights and 16-bit activations. Our AutoRound checkpoint ran on a frozen third-party patched vLLM image, with FP8 KV cache and additional RTX 3090 work. It is not a GGUF file that can be swapped into llama.cpp.

On one matched task, the complete W4A16/vLLM profile scored 100/100 in 502.14 seconds. Dynamic V3 Q4_K_M on llama.cpp scored 98/100 in 1,468.76 seconds. Server, cache, speculation and context limit also differed. The result selects an operating profile; it does not isolate weight quality.

Reading vendor accuracy claims

Unsloth describes Dynamic V3 as delivering about 10% more accuracy at the same size. That is the quantization vendor's claim under its evaluation method. We did not observe a 10% agent-quality increase.

Supporting that stronger statement would require:

1. identical prompts and runtime;
2. several weight formats under matched settings;
3. repeated runs with a frozen sampler;
4. a task set representing the process;
5. independent answer scoring.

We did not need that claim to select the experiment profile. Quality was at least comparable and the whole-task time margin was large. That is a narrower, defensible conclusion.

KV cache can consume the saving

Long context stores keys and values for prior tokens. It can occupy several gigabytes for one user and grows with concurrent sequences. A 16.5 GB weight file does not leave a clean 7.5 GB budget on a 24 GB card.

Reducing cache from Q8 to Q4 increases capacity but may affect access to long material. Test it with facts at different positions and a synthesis task. A successful server launch alone proves too little.

A selection procedure

1. Begin with the highest precision that fits the typical context.
2. Run the acceptance set and retain outputs.
3. Reduce weight or cache precision separately where possible.
4. Compare quality, time, VRAM and correction cost.
5. Set a required VRAM margin and reject profiles that repeatedly cross it.
6. Pin the winning artifact hash.

A document app may reveal loss in numbers and citations. A coding agent may reveal it in scope, contracts and tool syntax. One average perplexity number cannot replace both.

Syntalith performs this comparison in an AI process audit or as part of an AI app. The client receives the selected variant, rejected alternatives, measurements and an upgrade gate. Book a free scan when it is not yet clear that the process needs a 27B model.