Tess-4.5-27B GGUF

Tess-4.5-27B GGUF

GGUF builds of Tess-4.5, a 27B open-weights coding and agent model, for llama.cpp, with a vision projector and optional draft heads for faster generation. They suit people who run models locally.

Made by migtissera on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Tess-4.5 is an open-weights coding and agent model with 27 billion parameters, and this repository packs it as GGUF files for llama.cpp. It also holds a vision projector for image input and small multi-token-prediction draft heads that speed up generation. The author says it matches its base model, Qwen3.8-27B, on BenchLocal while using 25% fewer tokens.

Who it suits: It suits developers and hobbyists who want a coding assistant on their own machine, such as a single 24 GB GPU or a Mac with 32 GB or more.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Software | Developers | Run a terminal coding assistant locally | Point it at a project folder through the system prompt
Local AI | Hobbyists with a 24 GB GPU | Serve a quantized model with llama.cpp | Start llama-server and chat with it
Research | Inference tinkerers | Test speculative decoding with draft heads | Compare speed with and without the draft head

If you need full precision, you can skip these builds. The top quantization here is near-lossless but still quantized.

Where to use it: It runs in llama.cpp through llama-server. The card says the draft head feature needs a recent upstream llama.cpp build.

• llama.cpp on Windows, macOS or Linux
• A 24 GB GPU or a Mac with 32 GB or more, for the default build
• Machines with 48 GB or more, for the biggest build

Title: Jobs it does
Use case | Who | What they get back
Local coding agent | Developers | Replies that use the tools and files you provide
Image input | Multimodal users | Text answers about a picture, with the projector loaded
Faster generation | Power users | Higher speed from a draft head with the same model

How to set it up: You need a recent llama.cpp and enough memory for the build you pick. The card names Q4_K_M as the default for one 24 GB GPU.

• A recent upstream llama.cpp build
• A 24 GB GPU or a Mac with 32 GB or more for the default build
• Python with pip, if you fetch files with the Hugging Face tools

1. Install a recent llama.cpp.
2. Pick a build that fits your memory.
3. Start llama-server with the Hugging Face repository name, or with local files.

pip install -U huggingface_hub
hf download migtissera/Tess-4.5-27B-GGUF

How to use it: Start llama-server with the repository name and the speculative decoding option, which also fetches the matching draft head. Then chat with it. The card says Tess works best when you tell it where it is and pass your real tools.

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

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

You are Tess, an agentic coding assistant that operates in the user's terminal.
# Environment
Working directory: /path/to/your/project

For best results the card wants your actual tool definitions in the tools field, thinking left on, and sampling set to temperature 1.0, top_p 0.95 and top_k 20.

Title: Builds and files
Build | Size | Use it for
Q4_K_M | 16.5 GB | The default, for one 24 GB GPU or a Mac with 32 GB or more
Q6_K | 22.1 GB | More headroom and closer to full precision
Q8_0 | 28.6 GB | Near-lossless, for machines with 48 GB or more
Draft head Q8_0 | 3.2 GB | Faster generation with the same model
Draft head Q4_K_M | 2.0 GB | A smaller draft head for tight memory
Vision projector | 0.9 GB | Image input

Example: You start llama-server with the Q4_K_M repository name and the draft option, then write a system prompt that names your project folder. You ask it to find a bug in a file. It answers using the tools you passed.

For newcomers and experts: If you are new, this is a big model squeezed into files that run on a home computer. Install llama.cpp, start the server with the default build, and ask a short coding question.

If you know the field, the draft heads let llama.cpp guess several tokens ahead and keep the ones the main model agrees with, so the output is the same model's but faster. At extreme context lengths on a 24 GB GPU, the card suggests quantizing the KV caches or lowering the context size.

You start llama-server with a build
The draft head guesses ahead
The main model checks the guesses
Accepted tokens stream back faster

Good to know: The card measured speed on two machines with the Q4_K_M build, one request at a time, thinking on, in tokens per second.

Title: Tokens per second, Q4_K_M, code tasks
RTX 4090 without draft: 48.9 tok/s
RTX 4090 with draft head: 96.1 tok/s
Apple M4 Max without draft: 24.4 tok/s
Apple M4 Max with draft head: 37.8 tok/s

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the model files from Hugging Face, and your project files and tools when you connect them
• Not for: full precision, or an older llama.cpp build when you want the draft heads

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: gguf
• Parameters: 26.9B
• Weights: about 13.4 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 1,095,620
• Likes: 65
• License: apache-2.0
• Languages: en
• Based on: migtissera/Tess-4.5-27B
• Tags: gguf, tess, llama.cpp, agentic, coding, reasoning, image-text-to-text, conversational

Links

• Model: https://huggingface.co/migtissera/Tess-4.5-27B-GGUF
• Files: https://huggingface.co/migtissera/Tess-4.5-27B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download migtissera/Tess-4.5-27B-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.