Quantization by component
- All attention (full self_attn + GDN linear attention) — 4-bit float weights (RFA): IQ4_NL non-linear grid, group size 16, asymmetric, Hadamard-16 rotation, block-float scales stored as int8 mantissa + int8 exponent.
- MLP layers — same RFA 4-bit scheme.
- MTP speculative-decode head — its attention projections are RFA 4-bit too; its fc and MLP stay bf16.
- Kept in bf16 (not quantized) — vision encoder, GDN
in_proj_a/in_proj_b, embeddings, norms, and the lm_head.
Result: 20.5 GB (vs 55.6 GB bf16, 28.9 GB RFI) at essentially zero measured quality cost.
Serving context — 512K via YaRN
All evaluation below was run while serving at --max-model-len 524288 (512K tokens),
extended from the native 256K window with YaRN via --hf-overrides:
{"text_config": {"rope_parameters": {"rope_type": "yarn", "factor": 2.0,
"original_max_position_embeddings": 262144, "mrope_interleaved": true,
"mrope_section": [11, 11, 10], "partial_rotary_factor": 0.25,
"rope_theta": 10000000}}}
Evaluation results
Six builds measured with identical methodology, each against its own live vLLM endpoint (July 2026):
Qwen3.6-27B (bf16 base) → Tess-4-27B (the tune, bf16) → Tess-27B-RFI (int8+4-bit hybrid) →
Tess-27B-RFA (this build, all-4-bit) → Tess-FP8 (W8A8 block-128 FP8 sibling),
with Qwen3.6-35B-A3B (MoE, bf16) as a comparative reference.
Bold marks the best score in each row (ties all bolded).
Quantization cost (all-RFA 4-bit) — Tess-4-27B → Tess-27B-RFA
Table with columns: Metric, Change| Metric | Change |
|---|
| Checkpoint size | 55.6 → 20.5 GB (−63%) |
| WikiText-2 perplexity | 6.669 → 6.629 (−0.6%, better) |
| Codeneedle overall recall | 97.7% → 98.1% (best of family) |
| MC accuracy (4 tasks) | ≈ flat (−0.6 to +1.3 pp) |
| Tool-eval / GSM8K / MMLU / IFEval | 86 / 98% / 80% / 95% |
| MTP acceptance length | ~2.9 → ~2.8 (≈ flat) |
| Decode @ conc 1 (ISL 128) | 59.9 → 67.9 tok/s (+13%) |
| Decode @ conc 50 (ISL 128) | 528 → 423 tok/s (−20%) |
Quality
Table with columns: Metric, Qwen3.6-27B (base), Tess-4-27B, Tess-27B-RFI, Tess-27B-RFA, Tess-FP8, Qwen3.6-35B-A3B| Metric | Qwen3.6-27B (base) | Tess-4-27B | Tess-27B-RFI | Tess-27B-RFA | Tess-FP8 | Qwen3.6-35B-A3B |
|---|
| WikiText-2 PPL (n_ctx 2048, lower is better) | 7.0559 | 6.6691 | 6.6632 | 6.6292 | 6.6627 | 6.5092 |
| ARC-Challenge (acc_norm) | 59.30% | 60.84% | 60.41% | 60.32% |
Multiple-choice accuracy is lm-eval loglikelihood scoring, 0-shot.
Long-context positional recall (codeneedle)
Verbatim function recall under 10K–80K-token contexts.
Table with columns: Corpus, Qwen3.6-27B (base), Tess-4-27B, Tess-27B-RFI, Tess-27B-RFA, Tess-FP8, Qwen3.6-35B-A3B| Corpus | Qwen3.6-27B (base) | Tess-4-27B | Tess-27B-RFI | Tess-27B-RFA | Tess-FP8 | Qwen3.6-35B-A3B |
|---|
| Python | 100% | 100% | 100% | 100% | 99.55% | 99.09% |
| C++ | 98.12% | 98.12% | 98.44% | |
Table with columns: Bench, Qwen3.6-27B (base), Tess-4-27B, Tess-27B-RFI, Tess-27B-RFA, Tess-FP8, Qwen3.6-35B-A3B| Bench | Qwen3.6-27B (base) | Tess-4-27B | Tess-27B-RFI | Tess-27B-RFA | Tess-FP8 | Qwen3.6-35B-A3B |
|---|
| tool-eval final (full 69, TC-61 excl) | 86 | 85 | 87 | 86 | 87 | 90 |
| GSM8K (50q) | 98.0% | 94.0% | 98.0% | |
Decode throughput — tok/s output (ISL 128 / ISL 512)
vllm bench serve, random dataset, OSL 128, saturation, 4× R9700 (gfx1201), TP 4.
Table with columns: Concurrency, Qwen3.6-27B (base), Tess-4-27B, Tess-27B-RFI, Tess-27B-RFA, Tess-FP8, Qwen3.6-35B-A3B| Concurrency | Qwen3.6-27B (base) | Tess-4-27B | Tess-27B-RFI | Tess-27B-RFA | Tess-FP8 | Qwen3.6-35B-A3B |
|---|
| 1 | 57.0 / 61.4 | 59.9 / 63.8 | 75.3 / 69.3 | 67.9 / 58.2 | 86.0 / 85.1 | 91.9 / 114.4 |
| 10 | 304.5 / 233.6 | 289.6 / 227.8 | 281.5 / 219.2 | 292.2 / 194.8 |
MTP draft acceptance by work category
Measured from live serving logs, k=5 draft tokens, drafted-token-weighted aggregation.
Table with columns: Work category, Overall acceptance, Pos 1, Pos 2, Pos 3, Pos 4, Pos 5| Work category | Overall acceptance | Pos 1 | Pos 2 | Pos 3 | Pos 4 | Pos 5 |
|---|
| JSON generation | 87.3% | 96.4% | 92.2% | 86.8% | 82.9% | 78.0% |
| Math | 78.2% | 95.6% | 87.7% | 78.6% | 68.8% |
Notes
All builds serve on the tcclaviger/vllm:latest image, which has kernel tunes baked in.
TunableOp is untuned — GEMMs run on default heuristic-determined values.
Base Qwen3.6-27B figures are the 2026-07-12 re-measurement on the same tcclaviger/vllm:latest image and thinking-OFF methodology as every other build, replacing an earlier non-comparable run.
Credits
-
Tess-4-27B by Migel Tissera (migtissera/Tess-4-27B) — the model quantized here:
@misc{tissera2026tess4,
title = {Tess-4-27B},
author = {Migel Tissera},
year = {2026},
howpublished = {\url{https://huggingface.co/migtissera/Tess-4-27B}}
}
-
codeneedle (positional recall) originally by Alexander Ziskind, expanded test suite by tcclaviger (Rob Smith).
-
Tool-calling scenarios (incl. TC-61) run on tool-eval-bench by SeraphimSerapis (Tim Messerschmidt), scenario methodology adapted from ToolCall-15 by stevibe.
-
GSM8K, MMLU, and IFEval run via tool-eval-bench's built-in accuracy benchmarks at their defaults: GSM8K 8-shot CoT, MMLU 5-shot, IFEval zero-shot.