Tess-4-35B-A3B

Tess-4-35B-A3B (mirror + README)

This repository is an unmodified weight copy of the original model.
Original: migtissera/Tess-4-35B-A3B by Migel Tissera.
No fine-tuning, quantization, or weight changes were applied here. The only addition is this README — the upstream BF16 model card was empty, so documentation was assembled from the author’s GGUF card, the sibling Tess-4-27B card, config.json, and public announcements. Prefer the original repo and author sources for the canonical release.

---

Reasoning that scales with the problem — on an architecture light enough to run interactively on a single consumer machine.

Tess-4-35B-A3B is the Mixture-of-Experts member of the Tess-4 family by Migel Tissera. It is an agentic, thinking-native post-train of Qwen/Qwen3.6-35B-A3B: private deliberation in explicit <think>…</think> blocks, native tool use, and long-context agentic behavior — with sparse activation so you get ~35B-class breadth at ~3B-class decode cost.

---

The Tess-4 family

Tess-4 models are post-trained on 64K-token long-context agentic traces — real engineering work done with Fable-5, not synthetic single-turn data. Reasoning style is distilled from a three-model teacher ensemble (Opus-4.8, GPT-5.5, and GLM-5.2) into coherent deliberation patterns, using explicit <think>…</think> blocks for private reasoning before the visible answer.

The same dataset and recipe produced Tess-4-27B (dense), which community BenchLocal runs placed first at 81% (122/150) ahead of larger models including Qwen3.6-35B-A3B and Gemma-4-31B (community-reported; see Tess-4-27B discussions).

Tess-4-35B-A3B applies that recipe to the MoE base: 35B-class capacity with sparse activation for local/interactive use.

Why Tess-4 is different

• Weight-scaled reasoning — keeps routine steps tight; pours deliberation into planning, debugging, synthesis, and judgment calls.
• Agentic by design — native/parallel tool use and multi-step problem solving over real codebases and docs.
• Long-context post-train — 64K-token agentic traces; base architecture supports up to 262K positions.
• Multimodal — Qwen3.6 vision path retained in the checkpoint (vision_config, image/video token IDs).
• Honest, not sycophantic — trained toward grounded pushback over flattery (family claim from Tess-4-27B card).

---

Architecture (from config.json)

Sparse activation is the practical win on local hardware: ~3B active parameters per token with a wide expert pool.

---

Prompt format & thinking

Qwen3.5-family chat template with explicit think blocks (embedded as chat_template.jinja in this repo):

<|im_start|>user
Your prompt here<|im_end|>
<|im_start|>assistant
<think>
… private reasoning …
</think>
… visible answer …<|im_end|>

• transformers: processor.apply_chat_template(messages, add_generation_prompt=True)
• llama.cpp: --jinja (template is in the GGUFs / can be loaded from this repo’s jinja file)

Tool calling is intended to work through the same chat template path (family + GGUF card).

---

Available formats

GGUF + MTP (summary)

The GGUF repo is not a plain trunk dump: each main file embeds a multi-token-prediction (MTP) block for native self-speculative decoding (no separate -md draft model). Author-reported gains on Apple M4 Max (128 GiB), production gate p_min=0.70, include roughly +23–50% decode on code-heavy loads (e.g. Q8_0 code ~75 → ~110 tok/s). Every draft token is verified by the full trunk.

Important runtime notes (from the GGUF card):

• Built/validated on the opt-minimax branch of a trinity-cloud llama.cpp fork (--spec-type draft-mtp). Untested on stock llama.cpp; extra blk.40.* tensors may be rejected by strict loaders.
• Use -np 1 for speculation.
• Always set --spec-draft-p-min 0.70 (default 0.0 ungated drafting hurts throughput and greedy reproducibility).
• For strict eval harness determinism, serve without speculation flags (rare near-tie argmax flips possible under long greedy + MTP, worse on coarser quants).

Example (GGUF / fork):

llama-server -m Tess-4-35B-A3B-Q8_0-MTP-Q4D-Q4H-PV64K.gguf \
  -c 32768 -ub 2048 -np 1 -ngl 99 -fa on --jinja --metrics \
  --spec-type draft-mtp \
  --spec-draft-n-max 5 --spec-draft-n-min 0 --spec-draft-p-min 0.70 \
  --host 127.0.0.1 --port 8787

---

Quickstart

transformers (BF16)

Requires a recent transformers with Qwen3.5 / Qwen3.6 MoE multimodal support.

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "migtissera/Tess-4-35B-A3B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "user",
        "content": "Refactor this function and explain your reasoning.",
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

out = model.generate(**inputs, max_new_tokens=1024)
print(processor.decode(out[0][inputs["input_ids"].shape[-1] :], skip_special_tokens=True))

Multimodal: pass image (+ text) content blocks the same way as other Qwen3.5-VL / Qwen3.6 multimodal checkpoints (see Tess-4-27B card for a worked image example).

vLLM / SGLang

# vLLM (version must support qwen3_5_moe)
vllm serve migtissera/Tess-4-35B-A3B

# SGLang
python3 -m sglang.launch_server \
  --model-path migtissera/Tess-4-35B-A3B \
  --host 0.0.0.0 \
  --port 30000

Pin stack versions that explicitly list Qwen3.6-35B-A3B / qwen3_5_moe support; older builds will fail on architecture or hybrid attention.

llama.cpp (GGUF)

Prefer the GGUF repo and a runtime that understands the architecture (and MTP morph if you want speculation). For plain trunk decode, serve without --spec-type.

# Example: HF-hub launch of a quant tag (app/runtime dependent)
llama-server -hf migtissera/Tess-4-35B-A3B-GGUF:Q4_K_M --jinja -c 32768 -ngl 99

---

What it’s good at

• Agentic coding — explore unfamiliar repos, plan changes, run multi-step tool loops.
• Long-context technical work — large codebases and docs with relatively cheap KV on this hybrid stack.
• Judgment calls — structured analysis with evidence-backed pushback rather than default agreement.
• Interactive local use — MoE sparsity targets single-machine decode; MTP GGUFs push decode further on supporting runtimes.

---

Hardware ballpark

Context is comparatively cheap: author note ~20 KB/token scale when raising server context on this architecture (only a minority of layers keep full KV). Keep server -c ≥ client ceiling.

---

Related links

Code samples above that load migtissera/Tess-4-35B-A3B work unchanged against this mirror if you swap in this repo id.

---

Credits

• Model weights & Tess-4 training: Migel Tissera — original release migtissera/Tess-4-35B-A3B.
• Base architecture: Qwen/Qwen3.6-35B-A3B by the Qwen team.
• This repository: Unofficial mirror. Weights are a straight copy of the original; the only local addition is this README (upstream BF16 card was empty). Not affiliated with or endorsed by the author unless stated otherwise.

---

License

Apache License 2.0, inherited from the original model / Qwen base. Redistribution here does not change the license terms.

---

Citation

The original author release:

@misc{tess4-35b-a3b,
  title   = {Tess-4-35B-A3B: an agentic, thinking-native Mixture-of-Experts
             model with native MTP self-speculation},
  author  = {Tissera, Migel},
  year    = {2026},
  url     = {https://huggingface.co/migtissera/Tess-4-35B-A3B}
}

---

Unofficial mirror of Tess-4-35B-A3B (Migel Tissera) with documentation only. Community evaluations for the MoE variant are still catching up to the denser Tess-4-27B BenchLocal results; treat third-party scores as provisional until more independent runs land.