Tess-4-27B-GGUF

Tess-4-27B · GGUF

GGUF quantizations of migtissera/Tess-4-27B — an agentic, reasoning-native 27B model that scales its deliberation to the difficulty of the task: tight on the easy calls, deep on the hard ones. Built on Qwen/Qwen3.6-27B.

👉 Full model card, capabilities, and details: migtissera/Tess-4-27B

Quants

Run it

Download a quant (add the vision projector if you want image input):

hf download migtissera/Tess-4-27B-GGUF \
  Tess-4-27B-Q6_K.gguf mmproj-Tess-4-27B-F16.gguf \
  --local-dir ./tess-4-27b

llama.cpp

# text
llama-cli -m Tess-4-27B-Q6_K.gguf --jinja \
  -p "Refactor this function and explain your reasoning."

# image input (multimodal)
llama-mtmd-cli -m Tess-4-27B-Q6_K.gguf \
  --mmproj mmproj-Tess-4-27B-F16.gguf \
  --image photo.png -p "What's in this image?"

LM Studio / Ollama — load any text quant. For image input in LM Studio, keep mmproj-Tess-4-27B-F16.gguf in the same folder as the model and it's auto-detected. Use a recent llama.cpp runtime — older builds won't recognize the qwen35 architecture.

Faster inference (MTP)

Tess-4 ships a Multi-Token Prediction (MTP) head for speculative decoding: the head proposes the next token and the main model verifies it in the same forward pass, so you emit more than one token per pass — typically ~1.5–2× faster decoding on high-acceptance content (code, structured output). It's lossless — the main model verifies every token, so the output is identical whether MTP is on or off.

llama.cpp — add --mtp and it discovers the mtp- head in this repo automatically:

llama-cli -hf migtissera/Tess-4-27B-GGUF:Q4_K_M --mtp \
  -p "Refactor this function and explain your reasoning."

For local files, select the head explicitly with -md (recent llama.cpp build):

llama-cli -m Tess-4-27B-Q4_K_M.gguf -md mtp-Tess-4-27B-Q8_0.gguf --mtp -p "..."

Pick the MTP quant to match your budget:

• mtp-Tess-4-27B-Q8_0.gguf (3.2 GB) — higher draft accuracy → higher acceptance → more speedup (recommended).
• mtp-Tess-4-27B-Q4_K_M.gguf (2.0 GB) — smaller, for tight VRAM.

Keep one mtp- head alongside the model (or choose it with -md) so auto-discovery is unambiguous. LM Studio / Ollama auto-detect a same-folder mtp- head, just like the vision projector.

Using vLLM / SGLang? Reach for the full-precision BF16 model instead — MTP is already baked into that checkpoint (mtp_num_hidden_layers: 1); just enable speculative decoding (num_speculative_tokens: 1). No separate file needed.

Prompt format

Tess-4 uses the Qwen3.5-family chat template with explicit <think> … </think> reasoning. Pass --jinja in llama.cpp (or apply the chat template in your client) so the model reasons privately before it answers.

Credits & license

Quantized from migtissera/Tess-4-27B, built on Qwen/Qwen3.6-27B by the Qwen team — full credit for an outstanding base model. Released under Apache 2.0.

Part of the Tess series by Migel Tissera.