Tess-4-9B-GGUF

Tess-4-9B · GGUF

GGUF builds of migtissera/Tess-4-9B — an agentic, thinking-native 9B model that scales its deliberation to the difficulty of the task: tight on routine work, deep on the hard calls. Built on Qwen/Qwen3.5-9B-Base and post-trained on the same 64K-token Fable-5 agentic reasoning corpus as Tess-4-27B.

👉 Full model card, benchmark chart, training story, and Transformers usage: migtissera/Tess-4-9B

Why run Tess-4-9B locally?

• 🧠 Serious reasoning in a practical footprint. The recommended Q4_K_M build is only 5.63 GB.
• 📈 Strong measured gains. Tess-4-9B reaches 79.4% MMLU, 88.0% GSM8K, and 73% GPQA Diamond with extended context. See the full evaluation details.
• 🛠️ Agentic by design. Built for coding, tool use, long technical work, and evidence-driven judgment.
• 📏 Long-context architecture. The GGUF metadata carries Qwen3.5's native 262,144-token context; Tess was post-trained on 64K-token agentic traces.
• 🔒 Runs on your hardware. Use it through llama.cpp, LM Studio, Ollama, or another modern GGUF runtime.

Files

Which quant should I choose?

• Start with Q4_K_M for laptops, compact desktops, and the easiest fully local setup.
• Choose Q6_K when you have a little more memory and want a stronger quality margin.
• Choose Q8_0 for near-full-precision behavior while remaining much smaller than F16.
• Choose F16 for conversion work, maximum fidelity, or hardware with ample memory.

Run it

Directly from Hugging Face with llama.cpp

Use a recent llama.cpp build with Qwen3.5 (<code>qwen35</code>) and Jinja chat-template support:

# Interactive generation
llama-cli \
  -hf migtissera/Tess-4-9B-GGUF:Q4_K_M \
  --jinja \
  -c 65536 \
  -p "Inspect this code, identify the root cause, and propose a verified fix."

# OpenAI-compatible local server
llama-server \
  -hf migtissera/Tess-4-9B-GGUF:Q4_K_M \
  --jinja \
  -c 65536 \
  --host 127.0.0.1 \
  --port 8080

The 65,536-token context above leaves room for substantial reasoning. Reduce <code>-c</code> if your machine is memory-constrained; increase it only when your runtime and hardware budget allow.

Download a specific quant

hf download migtissera/Tess-4-9B-GGUF \
  Tess-4-9B-Q4_K_M.gguf \
  --local-dir ./tess-4-9b

Then run the local file:

llama-cli \
  -m ./tess-4-9b/Tess-4-9B-Q4_K_M.gguf \
  --jinja \
  -c 65536 \
  -p "Plan and implement this change, then verify your work."

LM Studio and Ollama

Load any text quant using a current runtime with Qwen3.5 architecture support. The model's agentic Qwen3.5 chat template is embedded in every GGUF file.

This GGUF repository currently contains the text model only. It does not include a separate multimodal projector, so image input is not advertised for these files; use the full-precision repository for the inherited vision path.

Prompt format and thinking

Tess-4 uses the Qwen3.5-family agentic chat template with explicit <code><think> … </think></code> reasoning blocks. The correct template is embedded in each GGUF.

Pass <code>--jinja</code> in llama.cpp, or configure your client to use the embedded template. Do not substitute a generic prompt template: Qwen-family models can behave materially differently when the renderer is wrong.

Tess-4-9B is a heavy reasoner on difficult problems. Give demanding math, science, debugging, and agentic tasks a generous context and generation budget. In GPQA evaluation, some answers exhausted a 24K output cap and benefited from retrying at 48K.

Build information

• Architecture: <code>qwen35</code>
• Native context encoded in GGUF: 262,144 tokens
• Model class: 9B
• Chat template: embedded Qwen3.5 agentic Jinja template
• Source checkpoint: migtissera/Tess-4-9B

Credits and license

Quantized from migtissera/Tess-4-9B, built on Qwen/Qwen3.5-9B-Base by the Qwen team — full credit for an outstanding base model. Released under Apache 2.0.

Part of the Tess series by Migel Tissera.