Add Tess-4-27B to the catalog (llama.cpp dual, external MTP, 262K) (#661)
Qwen3.5-based dense 27B (migtissera Q4_K_M GGUF); arch qwen35-dense,
64 layers, standard GQA — confirmed from the GGUF header. First
EXTERNAL-MTP compose in the catalog: the nextn draft ships as a
separate mtp-*.gguf, engaged via --spec-draft-model + --spec-type
draft-mtp (vs Deckard's embedded head).
Live-validated on dual 3090 (2026-07-09): serves at 262K, decode
~52 narr / 68 code, verify-stress 8/8 (NIAH clean to 240,634 tok),
soak-continuous PASS, 8-pack 115/150 think-off / 118/150 think-on —
ties/edges the qwen3.6-27b dual-max (109) and leads the agentic packs
(hermes 15/20 vs 9, cli-40 25/40 vs 20). Ships as caveats: streaming
tool-calls + thinking-ON -> finish=length (thinking-off for tools).
- models/tess-4-27b.yml + drafters/tess-mtp-gguf.yml (spec_method mtp_gguf)
- add mtp_gguf to llama-cpp-mainline supported_drafters (live-validated,
additive — mainline serves external draft-mtp on b9246)
- compose dual/migtissera-q4km/mtp.yml + registry entry + DEFAULTS row
- bump catalog counts (registry 62, disk 63, models 12, drafters 12)
- BENCHMARKS.md: Tess dual-card section with dual-max comparison
Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm
Co-authored-by: noonghunna <10742901+noonghunna@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
8eff77e2be