tcclaviger/Tess-27B-RFA

[!IMPORTANT] ## All-RFA 4-bit quant of migtissera/Tess-4-27B Runtime: requires `tcclaviger/vllm:latest` — an RDNA 4 (gfx12xx) vLLM image and the only build with the RFA kernels; no other vLLM build loads these weights. Not validated on any other hardware at this time.

Tess-27B-RFA

Aggressive all-RFA 4-bit quantization of Tess-4-27B by Migel Tissera — an agentic, thinking-native finetune of Qwen3.6-27B. Unlike the sibling Tess-27B-RFI (int8 attention), this build runs every attention path at 4-bit: full attention, GDN linear attention, and even the MTP speculative head's attention. All credit for the model to its author; this repo only changes the numerics.

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

Quality

Multiple-choice accuracy is lm-eval loglikelihood scoring, 0-shot.

Long-context positional recall (codeneedle)

Verbatim function recall under 10K–80K-token contexts.

Tool calling & accuracy benches

Decode throughput — tok/s output (ISL 128 / ISL 512)

vllm bench serve, random dataset, OSL 128, saturation, 4× R9700 (gfx1201), TP 4.

MTP draft acceptance by work category

Measured from live serving logs, k=5 draft tokens, drafted-token-weighted aggregation.

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.