GGUF conversions of migtissera/Tess-4-27B-EAGLE3, the EAGLE-3 speculative-decoding draft head for migtissera/Tess-4-27B, for use with mainline llama.cpp's external EAGLE-3 path (--spec-type draft-eagle3).

File	Size	Notes
`Tess-4-27B-EAGLE3-Q4_K_M.gguf`	403 MB	Recommended — acceptance is flat across draft quants, so the smallest draft wins on latency
`Tess-4-27B-EAGLE3-Q8_0.gguf`	629 MB
`Tess-4-27B-EAGLE3-BF16.gguf`	1.22 GB	Direct converter output

Usage

bash

llama-server \
  -m Tess-4-27B-Q4_K_M.gguf \
  --spec-type draft-eagle3 \
  --spec-draft-model Tess-4-27B-EAGLE3-Q4_K_M.gguf \
  --spec-draft-n-max 2

--spec-draft-n-max 2 is the measured sweet spot on this head — deeper drafting is strictly worse (per-position acceptance decays; rejected drafts waste target verification). Speculative decoding is output-lossless: the target verifies every token, confirmed at benchmark granularity (150-scenario quality suite, within run-to-run noise of no-spec and of a DFlash draft).

Measured performance

Acceptance on this head is modest — it accelerates Tess-4, but a stronger draft beats it where one exists:

Rig	Config	Acceptance	Throughput
1× RTX 4090 (Q4_K_M target, ctx 98K)	n_max=2	49.1%	~76–78 code tok/s (vs 50.4 no-spec — 1.55×)
1× RTX 4090	n_max=4	27.4–28.0% (flat across BF16/Q8_0/Q4_K_M draft)	62–71 tok/s
2× RTX 3090 (club-3090, llama.cpp b9967)	n=2	40.8%	47.5 narr / 61.3 code tok/s

For context on the same rigs, ~0.70-acceptance heads (DFlash, external MTP) reach 1.9–2×. See noonghunna/club-3090#662 for forensics on this head's acceptance ceiling and #665 / PR #674 for the quality benchmarks behind these numbers. Notably, this head's acceptance is feed-robust: it scores in the same band fed hidden states (vLLM/SGLang, in-graph) or fed tokens (llama.cpp external path).

Conversion recipe

Converted at llama.cpp b9932 (a646006f0); independently reproduced at b9967 — the recipe is version-robust across that span.

bash

# The target dir needs ONLY the base model's config.json + tokenizer files
# (~12 MB download) — no 27B weights required.
python convert_hf_to_gguf.py <Tess-4-27B-EAGLE3-dir> \
  --target-model-dir <Tess-4-27B-metadata-dir>          # → BF16 GGUF
llama-quantize Tess-4-27B-EAGLE3-BF16.gguf Tess-4-27B-EAGLE3-Q4_K_M.gguf Q4_K_M
llama-quantize Tess-4-27B-EAGLE3-BF16.gguf Tess-4-27B-EAGLE3-Q8_0.gguf Q8_0

The converter reads the target's hidden size (5120) from its config, selects extraction layers [2, 32, 61], and writes the 32K→248K d2t vocab map into the GGUF (eagle3.target_layers / eagle3.target_hidden_size metadata, inspectable with llama-gguf).

License

Apache 2.0, matching the source head and Tess-4-27B. Draft head by Migel Tissera; GGUF conversion and llama.cpp benchmarks as credited above.

YourHighnessLA/Tess-4-27B-EAGLE3-GGUF

Tess-4-27B EAGLE-3 Draft — GGUF

Usage

Measured performance

Conversion recipe

License

YourHighnessLA/
Tess-4-27B-EAGLE3-GGUF