Qwen3.8-27B-ASCII-Condensed

Qwen3.8-27B with a condensed ASCII-only vocabulary and a matching DFlash2 draft model. Meant to be used together with troed/llama.cpp-adaptive-kv-streaming, a fork of Raymond's KV streaming.

• vocabulary reduced from 248,320 to 129,006 rows in the embedding and output head
• the freed VRAM becomes KV cache: 160K context on a 16 GB GPU
• no retraining: surviving weights are bit-identical to the sources below

Note: In v2 of Raymond's phase arena there's support for MTP but currently not DFlash2. You thus only need the model file if using my settings. The DFlash2 of course works fine with regular llama.cpp.

Files

File	Size	Role
Qwen3.8-27B-ASCII-Condensed-IQ4_XS-3.84bpw.gguf	11.4 GiB	target model
Qwen3.8-27B-ASCII-Condensed-DFlash2-Q2_K_S-MIX.gguf	511 MiB	DFlash2 draft model (load with md =)

How the files were made

Target

Created from byteshape/Qwen3.8-27B-GGUF (its Qwen3.8-27B-IQ4_XS-3.84bpw.gguf) using bsaleh03's ASCII-Condensed-prune-tools. The vocab rows were gathered directly in quantized space and the tokenizer was rewritten to match, so every surviving weight is bit-identical to the source: no dequantization, no requantization, no retraining.

Draft

Created from the original HermiHg/Qwen3.8-27B-DFlash2-Q2_K_S-MIX-GGUF, with the same condensation applied to the draft's vocab tensors and tokenizer. No weights were changed beyond the row subset.

Requirements

A build of troed/llama.cpp-adaptive-kv-streaming:

cmake -B build -DGGML_NATIVE=ON -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CUDA_FA_QUANTS=q8_0-q4_0 -DGGML_CUDA=ON
cmake --build build --config Release -j

The files are standard GGUFs and load in any recent llama.cpp, but the shared-device-memory-mib settting need the fork.

Usage

hf download troed/Qwen3.8-27B-ASCII-Condensed --local-dir models

Config for the fork's llama-server (local paths adjusted):

[Qwen3.8-27B]
spec-type = draft-mtp
spec-draft-n-max = 5
spec-draft-p-min = 0.8
device-draft = CUDA0
n-gpu-layers-draft = all
# If all 16GB are available to the model, else lower this value
shared-device-memory-mib = 3904
# Any suitable template
chat-template-file = chat_template_qwen3.8.jinja
m = Qwen3.8-27B-ASCII-Condensed-IQ4_XS-3.84bpw.gguf
ctx-size = 200192
n-gpu-layers = 99
batch-size = 256
ubatch-size = 256
cache-type-k = q8_0
cache-type-v = q4_0
fit = off
parallel = 1
temp = 1.0
top-p = 0.95
top-k = 20
min-p = 0.0
presence-penalty = 0.0
repeat-penalty = 1.0
reasoning = on
reasoning-preserve = on
# Original GGUF-converted mmproj
mmproj = Qwen3.8-mmproj-BF16.gguf
load-mode = none
flash-attn = on

chat-template-file is any Qwen 3.8 jinja chat template. No mmproj is provided here; use e.g. the one from byteshape/Qwen3.8-27B-GGUF (mmproj-bf16.gguf).

Performance

On an RTX 5060 Ti 16 GB + 96 GB DDR5, with the config above:

• prompt processing: ~500-1000 t/s
• token generation: ~30-50 t/s

Full setup walkthrough: 16 GB VRAM llama-server configs.

Language support

The vocabulary is ASCII only. The 256 byte-level fallback tokens are always kept, so non-ASCII text still decodes correctly, it just costs more tokens per character. If you need Latin-extended, Greek, currency or box-drawing characters, re-run the prune tools with a wider policy.

Credits

• Qwen for Qwen3.8-27B
• Raymond Huang for the KV cache streaming work
• ByteShape for the IQ4_XS-3.84bpw quantization (weights unaltered)
• bsaleh03 for the vocabulary pruning tools
• HermiHg for the Q2_K_S-MIX DFlash2 draft (weights unaltered)

Licensed under Apache-2.0, inherited from the base model.