Why Tess-4 is different
- 🧠 Weight-scaled reasoning. Tess-4 keeps routine steps tight and pours deliberation into the hard ones — planning, debugging, synthesis, judgment calls. It thinks proportionally to the difficulty of the moment.
- 🛠️ Agentic by design. Native tool use and disciplined multi-step problem solving. It reads a codebase, builds a real mental model, and acts on it.
- 📏 Long-context, trained at 64K. Post-trained on 64K-token long-context agentic traces, so it holds a large working set without losing the thread.
- ⚡ Efficient at scale. A 512-expert MoE with 12B active parameters on a mostly-Mamba backbone: fast decoding and a small KV cache for long sessions.
- 👁️ Multimodal. Inherits Nemotron's C-RADIO vision encoder — text and image in. (For GGUF, pair with the included vision projector.)
- 🤝 Honest, not sycophantic. Trained to give grounded, evidence-based pushback instead of flattery.
The reasoning traces
Tess-4's signature is how it thinks. The reasoning/thinking traces used to train it were a best-case approximation of Fable-5, produced by a combination of Opus-4.8, GPT-5.5, and GLM-5.2 working together as a team — a multi-model teacher ensemble distilled into a single, coherent reasoning style.
The result is a model that reasons prospectively — predicting, verifying, and weighing alternatives before acting — rather than narrating after the fact.
Tess-4-120B-A12B uses NVIDIA's Nemotron-3.5 chat template with <think> … </think> reasoning blocks:
<|im_start|>userYour prompt here<|im_end|><|im_start|>assistant<think>… the model's private reasoning …</think>… the model's answer …<|im_end|>
Keep the reasoning in history. Tess-4 was trained with the reasoning of every earlier assistant turn left in context, so this repo's chat template defaults to truncate_history_thinking=False (NVIDIA's template defaults to True). In agentic loops, send each assistant turn's reasoning_content back with the conversation. Tool calls use the Qwen3-Coder XML format (--tool-call-parser qwen3_coder in vLLM/SGLang).
Sampling: temperature=1.0, top_p=0.95 (set in generation_config.json).
This repo — full-precision weights:
Table with columns: Format, ~Size, Best for| Format | ~Size | Best for |
|---|
| BF16 safetensors | 249 GB | transformers · vLLM · SGLang |
GGUF quants → migtissera/Tess-4-120B-A12B-GGUF
Table with columns: Quant, ~Size, Best for| Quant | ~Size | Best for |
|---|
| Q4_K_M | 88.4 GB | 128 GB Macs and multi-GPU rigs — best quality/size |
| Q8_0 | 131.6 GB | effectively lossless |
| BF16 | 247.4 GB | full precision · source for custom quants |
| mmproj (F16) | 1.7 GB | vision projector, pairs with any text GGUF |
Quickstart
vLLM
vllm serve migtissera/Tess-4-120B-A12B \ --trust-remote-code --tensor-parallel-size 4 --max-model-len 65536 \ --reasoning-parser nemotron_v3 \ --enable-auto-tool-choice --tool-call-parser qwen3_coder
llama.cpp (GGUF)
hf download migtissera/Tess-4-120B-A12B-GGUF --include "Q4_K_M/*" --include "mmproj-*" --local-dir ./tess-4-120bllama-server -m ./tess-4-120b/Q4_K_M/Tess-4-120B-A12B-Q4_K_M-00001-of-00006.gguf --jinja -c 65536 --temp 1.0 --top-p 0.95
Needs a llama.cpp build with Nemotron-3.5 Super VL support (added in ggml-org/llama.cpp#30520, October 2026).
What it's good at
- Agentic coding — exploring unfamiliar repos, planning changes, and executing multi-step work with tools.
- Long-context work — reasoning over large codebases and documents without dropping context.
- Technical & product judgment — honest, structured analysis that pushes back with evidence rather than agreeing by default.
Notes
- Post-training ran at 64K context. The base model supports up to 1M tokens, but behavior beyond 64K has not been tested.
- The vision encoder and the built-in multi-token-prediction (MTP) head are NVIDIA's originals, unchanged by post-training. If you use MTP speculative decoding, expect lower draft acceptance than on the base model.
Credits
Tess-4-120B-A12B is built on NVIDIA Nemotron-3.5-Super-VL-120B-A12B by NVIDIA — full credit to the Nemotron team for an outstanding base model.
License
Use of this model is governed by the OpenMDW License Agreement, version 1.1, inherited from the base model. A copy is included in LICENSE.
Citation
@misc{tissera2026tess4_120b, title = {Tess-4-120B-A12B}, author = {Migel Tissera}, year = {2026}, howpublished = {\url{https://huggingface.co/migtissera/Tess-4-120B-A12B}}, note = {Built on nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16}}
Tess-4-120B-A12B — part of the Tess series by Migel Tissera. Evaluations forthcoming.