XpressAI/Qwen3.8-27B-RYS-GGUF

Qwen3.8-27B — RYS Layer Surgery (GGUF)

An experimental modification of Qwen3.8-27B, produced by RYS layer duplication: no training and no weight changes, just running language layers 16–19 (zero-based) a second time during the forward pass.

The source quant is unsloth/Qwen3.8-27B-GGUF, revision f1bfb127c64f7072bdd2cad55f258b9c8b2910fe. The method is based on David Ng's RYS work.

TL;DR

This variant is intended for low-effort reasoning mode. It improved a small reasoning probe by +11.76 percentage points and a fixed 1,006-case BFCL v4 single-turn evaluation by +0.60 pp micro accuracy. The BFCL gain is small and not statistically decisive (36 RYS-only wins versus 30 base-only wins, two-sided exact McNemar p = 0.539), so treat this as an experimental rather than universally superior release.

The larger EQ-140 result was effectively flat-to-slightly-down (−0.99). Math also regressed in the search probe; math was recorded as a diagnostic, not used as a release gate, because it is an expected and fine-tunable tradeoff here.

Evaluation	Base	RYS 16–19	Change
Reasoning probe (17 prompts)	58.82%	70.59%	+11.76 pp
EQ-140 (139 scored)	82.65	81.66	−0.99
BFCL micro accuracy (1,006 cases)	82.60%	83.20%	+0.60 pp
BFCL category macro mean	80.54%	81.81%	+1.28 pp
Math probe (diagnostic only)	1.000	0.750	−0.250

File

File	Quant	Language layers	Size	SHA-256
Qwen3.8-27B-rys_16-19-UD-Q4_K_XL.gguf	UD-Q4_K_XL	68	18.9 GB	a3013e582490a386d4b8fafbaad07bfec65e7727fb5dcf992914babdc2d65ed4

The terminal MTP block from the source GGUF is retained and shifted after the four inserted language layers.

BFCL results

The evaluation used a fixed 1,006-case manifest spanning 13 BFCL v4 single-turn categories. Categories with at least 100 entries used 100 cases; smaller categories used all available entries. This is not the current full 3,641-case BFCL suite and excludes multi-turn, memory, and web-search tasks.

Category	N	Base	RYS 16–19	Change
irrelevance	100	86.00%	88.00%	+2.00
multiple	100	94.00%	91.00%	−3.00
parallel	100	90.00%	92.00%	+2.00
parallel_multiple	100	82.00%	84.00%	+2.00
simple_java	100	54.00%	53.00%	−1.00
simple_javascript	50	64.00%	62.00%	−2.00
simple_python	100	94.00%	94.00%	0.00
live_irrelevance	100	95.00%	99.00%	+4.00
live_multiple	100	79.00%	77.00%	−2.00
live_parallel	16	87.50%	87.50%	0.00
live_parallel_multiple	24	62.50%	70.83%	+8.33
live_relevance	16	75.00%	81.25%	+6.25
live_simple	100	84.00%	84.00%	0.00
Micro accuracy	1,006	82.60%	83.20%	+0.60
Macro mean	13 categories	80.54%	81.81%	+1.28

Inference used llama.cpp's OpenAI-compatible chat-completions endpoint with native tool use, the embedded Unsloth chat template, four slots, Q8 KV cache, and these matched settings for both models:

reasoning_effort = low
enable_thinking  = true
preserve_thinking = true
temperature = 1.0
top_p       = 0.95
top_k       = 20
min_p       = 0.0
seed        = 3407
max_tokens  = 2048

The RYS pass generated in 59:16. Normalizing the clean 991-case portion of the resumed base pass to 1,006 cases gives about 59:41, so wall time was effectively flat in this run. RYS produced about 2.0% more output tokens.

What changed?

Qwen3.8-27B uses a repeating three-DeltaNet/one-full-attention hybrid. RYS insertions therefore need to preserve four-layer blocks. This model duplicates the half-open layer window [16, 20):

Base:  0 → … → 15 → 16 → 17 → 18 → 19 → 20 → … → 63 → MTP
RYS:   0 → … → 15 → 16 → 17 → 18 → 19
                     → 16 → 17 → 18 → 19 → 20 → … → 63 → MTP

The sweep covered every compatible four-layer window at stride four, followed by neighboring starts around the strongest reasoning regions. Layers 16–19 were selected because they gave the best reasoning-first result while keeping the small EQ change near zero. The full BFCL result points in the same positive direction, though modestly.

Usage

Use a recent llama.cpp build. Start the server with thinking support and the embedded Jinja template:

./build/bin/llama-server \
  -m Qwen3.8-27B-rys_16-19-UD-Q4_K_XL.gguf \
  --port 8081 -np 4 -ngl all \
  -b 2048 -ub 1024 --flash-attn on \
  --jinja --no-host --reasoning on --reasoning-preserve \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0

Request low-effort reasoning explicitly:

{
  "model": "qwen3.8-27b-rys",
  "messages": [{"role": "user", "content": "Your prompt"}],
  "temperature": 1.0,
  "top_p": 0.95,
  "top_k": 20,
  "min_p": 0.0,
  "chat_template_kwargs": {
    "enable_thinking": true,
    "reasoning_effort": "low",
    "preserve_thinking": true
  }
}

The embedded Unsloth template already supports low-effort and reasoning-off modes. A quick comparison with froggeric/Qwen-Fixed-Chat-Templates v22 found no material reasoning-quality difference, so this GGUF retains the source template.

When to use this

• You run Qwen3.8 with low-effort reasoning enabled.
• You value the directional gains on parallel/relevance tool-use tasks.
• You are comfortable evaluating an experimental layer-surgery variant for your own workload.

When not to use this

• You need a proven across-the-board improvement; the aggregate BFCL gain is small and not statistically decisive.
• EQ or math performance must not regress at all.
• Your workload is dominated by multiple, live_multiple, or simple JavaScript function calling, where the base model did slightly better here.

Credits

• Qwen for Qwen3.8-27B
• Unsloth for the source UD-Q4_K_XL quant
• David Ng for the RYS method
• llama.cpp for local inference
• Berkeley Function-Call Leaderboard for the evaluation harness

License

Apache 2.0, inherited from Qwen3.8-27B and the source GGUF.