Tess-4.5 · GGUF

October 2026 · Model card · Continuity Research

The frontier, at home.

Tess-4.5 is a 27-billion-parameter open-weights model for real engineering work. It reasons through every step, works with the tools and files actually in front of it, and runs on a single GPU or a laptop. On BenchLocal it matches its base, Qwen3.8-27B, with 25% fewer tokens. This repository holds the GGUF quantizations for llama.cpp, the vision projector and the multi-token-prediction draft heads.

For benchmarks and the full story, see the model card.

Files

File	Size	Use it for
`Tess-4.5-27B-Q4_K_M.gguf`	16.5 GB	One 24 GB GPU, or a 32 GB+ Mac. The default.
`Tess-4.5-27B-Q6_K.gguf`	22.1 GB	More headroom, closer to full precision.
`Tess-4.5-27B-Q8_0.gguf`	28.6 GB	Near-lossless, for 48 GB+ machines.
`mtp-Tess-4.5-27B-Q8_0.gguf`	3.2 GB	MTP draft head. Up to 2× faster generation, same model.
`mtp-Tess-4.5-27B-Q4_K_M.gguf`	2.0 GB	Smaller MTP draft head for tight memory.
`mmproj-Tess-4.5-27B-F16.gguf`	0.9 GB	Vision projector, for image input.

Run it

llama-server -hf migtissera/Tess-4.5-27B-GGUF:Q4_K_M --spec-type draft-mtp --jinja -ngl 99 -c 65536

With -hf, --spec-type draft-mtp also fetches the matching mtp- draft head. With local files:

llama-server -m Tess-4.5-27B-Q4_K_M.gguf -md mtp-Tess-4.5-27B-Q8_0.gguf \
  --spec-type draft-mtp --jinja -ngl 99 -ngld 99 -c 65536

MTP speculative decoding needs a recent upstream llama.cpp build. On a 24 GB GPU at very long contexts, quantize the KV caches (--cache-type-k q4_0 --cache-type-v q4_0, and -ctkd q4_0 -ctvd q4_0 for the draft) or lower -c.

Speed

Q4_K_M, upstream llama.cpp, one request at a time, thinking on.

Machine	Without draft	With MTP draft, code	With MTP draft, shell
NVIDIA RTX 4090	48.9 tok/s	96.1 tok/s	85.5 tok/s
Apple M4 Max	24.4 tok/s	37.8 tok/s	33.1 tok/s

Give Tess its bearings

Tess works best when it knows where it is. State the working directory in the system prompt and attach your real tool definitions:

You are Tess, an agentic coding assistant that operates in the user's terminal.

# Environment

Working directory: /path/to/your/project

Pass your real tools in the tools field, keep thinking on, and sample with temperature=1.0, top_p=0.95, top_k=20.

Apache 2.0. Built on Qwen3.8-27B. Thank you to the Qwen team and the llama.cpp community.