Qwen3.5-27B-TQ3_4S

Clean base TQ3_4S GGUF release for Qwen3.5-27B.

TQ3_4S is a 3.5-bit Walsh-Hadamard-transform weight format with four per-8 scales per 32-weight block.

Summary

• Format: TQ3_4S
• Model size: about 12.9 GiB
• Target runtime: public TurboQuant-enabled llama.cpp
• Intended use: local inference on consumer GPUs
• Multimodal projector included: mmproj-BF16.gguf

Quality

Qwen3.5-27B, wiki.test.raw, c=2048:

Notes:

• TQ3_4S and Q3_K_S are full-pass llama-perplexity results.
• TQ3_1S is also a full-pass llama-perplexity result at c=2048.
• EXL3 3.0bpw is from a local 145 x 2048 eval, not llama-perplexity.
• This 27B result should not be read as evidence that plain TQ3_4S works equally well on smaller dense models.

Runtime

This model requires the public TurboQuant runtime fork:

• https://github.com/turbo-tan/llama.cpp-tq3
• LM Studio setup: docs/backend/LMStudio.md

Build and run:

git clone https://github.com/turbo-tan/llama.cpp-tq3.git
cd llama.cpp-tq3

cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

./build/bin/llama-server \
  -m /path/to/Qwen_Qwen3.5-27B-TQ3_4S.gguf \
  -ngl 99 \
  -fa on \
  -c 8192 \
  -ctk q8_0 -ctv q8_0 \
  --cache-ram 0 \
  --no-warmup --jinja \
  --reasoning off --reasoning-budget 0 --reasoning-format deepseek \
  --port 8090

Vision / Image Input

Use the included projector:

./build/bin/llama-server \
  -m /path/to/Qwen_Qwen3.5-27B-TQ3_4S.gguf \
  -mm /path/to/mmproj-BF16.gguf \
  -ngl 99 -c 8192 -np 1 \
  -ctk q8_0 -ctv q8_0 -fa on \
  --cache-ram 0 --no-warmup --jinja \
  --reasoning off --reasoning-budget 0 --reasoning-format deepseek \
  --no-mmproj-offload

If your frontend says image input is unsupported, it is usually still pointed at a server instance that was started without --mmproj.

Notes

This upload is the clean base TQ3_4S release, not the private KLD-guided mixed-precision variants.

Credits

• llama.cpp
• Qwen3.5-27B
• Walsh-Hadamard / transform-quantization line including RaBitQ, TurboQuant, and related work