Tess-4-27B-abliterated API & Inference Endpoint

Abliteration parameters

Table with columns: Parameter, Value
Parameter	Value
direction_index	31.66
attn.o_proj.max_weight	1.33
attn.o_proj.max_weight_position	58.63
attn.o_proj.min_weight	0.00
attn.o_proj.min_weight_distance	36.46
mlp.down_proj.max_weight	1.49
mlp.down_proj.max_weight_position	39.82
mlp.down_proj.min_weight	1.33
mlp.down_proj.min_weight_distance	36.43

Performance

Table with columns: Metric, This model, Original model (migtissera/Tess-4-27B)
Metric	This model	Original model (migtissera/Tess-4-27B)
KL divergence	0.0006	0 (by definition)
Refusals	81/100	90/100

Tess-4-27B

Reasoning that scales with the problem. An agentic, thinking-native model that deliberates harder exactly when it matters — and gets out of its own way when it doesn't.

Tess-4-27B is the first Tess release in two years, and the first that reasons. Built on Qwen/Qwen3.6-27B by Migel Tissera, it's post-trained on a deliberate blend: 64K-token long-context agentic traces — real engineering work done with Fable-5, not synthetic generations — with a reasoning style approximated from Fable-5 by a three-model teacher ensemble (Opus-4.8, GPT-5.5, and GLM-5.2) fused into one coherent voice.

The result is a 27B model that thinks like a senior engineer: form a hypothesis, act, verify, and reason with real density on the turns that actually deserve it — not a model that narrates its way to an answer it already had.

Community Performed Benchmarks

Currently best-in-class for BenchLocal

Table with columns: Rank, Model, Score, Result
Rank	Model	Score	Result
1	Tess-4-27B (Q8)	81%	122/150
2	Qwen3.6-35B-A3B (UD-Q8_K_XL)	78%	117/150
3	Gemma-4-31B (Q6 · 180k ctx)	78%	117/150
4	Qwopus3.6-27B Coder-Compat (Q6_K)	77%	116/150

References

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 doesn't ramble; it thinks proportionally to the difficulty of the moment.
🛠️ Agentic by design. Native, parallel 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.
👁️ Multimodal. Inherits Qwen3.6's vision tower — 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.

Prompt format & thinking

Tess-4 uses the Qwen3.5-family chat template with explicit <think> … </think> reasoning blocks. The model reasons privately, then produces its visible answer:

markdown
<|im_start|>user
Your prompt here<|im_end|>
<|im_start|>assistant
<think>
… the model's private reasoning …
</think>

… the model's answer …<|im_end|>

Apply it automatically via tokenizer.apply_chat_template(messages, add_generation_prompt=True), or --jinja in llama.cpp.

Available formats

This repo — full-precision weights:

Table with columns: Format, ~Size, Best for
Format	~Size	Best for
BF16 safetensors	52 GB	transformers · vLLM · SGLang

GGUF quants → migtissera/Tess-4-27B-GGUF

Table with columns: File, Format, ~Size, Best for
File	Format	~Size	Best for
`Tess-4-27B-Q4_K_M.gguf`	Q4_K_M	16.5 GB	smallest — great quality/size · most popular
`Tess-4-27B-Q6_K.gguf`	Q6_K	22 GB	near-lossless
`Tess-4-27B-Q8_0.gguf`	Q8_0	28 GB	effectively lossless

Faster inference

⚡ Tess-4-27B-EAGLE3 — a speculative-decoding draft trained on Tess-4's own outputs: 1.76× average decode speedup, up to 2.4× on reasoning (measured on H100; lossless — outputs are identical). SGLang: --speculative-algorithm EAGLE3 --speculative-draft-model-path migtissera/Tess-4-27B-EAGLE3; vLLM: --speculative-config '{"method":"eagle3","model":"migtissera/Tess-4-27B-EAGLE3","num_speculative_tokens":4}'.
🧮 Tess-4-27B-NVFP4 — 4-bit NVFP4 (19 GB, −63%), Blackwell-native W4A4, calibrated on Tess-4's own generations. Quantization and speculative decoding stack.

Quickstart

llama.cpp / LM Studio (GGUF)

Grab the quant(s) from migtissera/Tess-4-27B-GGUF:

bash
hf download migtissera/Tess-4-27B-GGUF \
  Tess-4-27B-Q4_K_M.gguf mmproj-Tess-4-27B-F16.gguf \
  --local-dir ./tess-4-27b

bash
# text
llama-cli -m Tess-4-27B-Q4_K_M.gguf --jinja -p "Refactor this function and explain your reasoning."

# with images (multimodal)
llama-mtmd-cli -m Tess-4-27B-Q4_K_M.gguf \
  --mmproj mmproj-Tess-4-27B-F16.gguf \
  --image photo.png -p "What's in this image?"

LM Studio: put mmproj-Tess-4-27B-F16.gguf in the same folder as the model file — LM Studio auto-detects it and enables image input. (Use a recent runtime; older llama.cpp builds won't recognize the architecture.)

transformers

python
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id = "migtissera/Tess-4-27B"
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": "Explain the tradeoffs of LoRA vs full fine-tuning."}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(inputs, max_new_tokens=1024)
print(processor.decode(out[0], skip_special_tokens=True))

(Requires a recent transformers with Qwen3.5/3.6 support.)

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.

Credits

Tess-4-27B is built on Qwen/Qwen3.6-27B by the Qwen team — full credit to them for an outstanding base model. Tess-4 inherits its Qwen3.5-family vision-language architecture and its Apache 2.0 license.

License

Released under the Apache License 2.0, inherited from the base model. See LICENSE.

Citation

bibtex
@misc{tissera2026tess4,
  title        = {Tess-4-27B},
  author       = {Migel Tissera},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/migtissera/Tess-4-27B}},
  note         = {Built on Qwen/Qwen3.6-27B}
}

Tess-4-27B — part of the Tess series by Migel Tissera. Evaluations forthcoming.

Abliteration parameters

Table with columns: Parameter, Value
Parameter	Value
direction_index	31.66
attn.o_proj.max_weight	1.33
attn.o_proj.max_weight_position	58.63
attn.o_proj.min_weight	0.00
attn.o_proj.min_weight_distance	36.46
mlp.down_proj.max_weight	1.49
mlp.down_proj.max_weight_position	39.82
mlp.down_proj.min_weight	1.33
mlp.down_proj.min_weight_distance	36.43

Performance

Table with columns: Metric, This model, Original model (migtissera/Tess-4-27B)
Metric	This model	Original model (migtissera/Tess-4-27B)
KL divergence	0.0006	0 (by definition)
Refusals	81/100	90/100

Tess-4-27B

Reasoning that scales with the problem. An agentic, thinking-native model that deliberates harder exactly when it matters — and gets out of its own way when it doesn't.

Community Performed Benchmarks

Currently best-in-class for BenchLocal

Table with columns: Rank, Model, Score, Result
Rank	Model	Score	Result
1	Tess-4-27B (Q8)	81%	122/150
2	Qwen3.6-35B-A3B (UD-Q8_K_XL)	78%	117/150
3	Gemma-4-31B (Q6 · 180k ctx)	78%	117/150
4	Qwopus3.6-27B Coder-Compat (Q6_K)	77%	116/150

References

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 doesn't ramble; it thinks proportionally to the difficulty of the moment.
🛠️ Agentic by design. Native, parallel 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.
👁️ Multimodal. Inherits Qwen3.6's vision tower — 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

The result is a model that reasons prospectively — predicting, verifying, and weighing alternatives before acting — rather than narrating after the fact.

Prompt format & thinking

Tess-4 uses the Qwen3.5-family chat template with explicit <think> … </think> reasoning blocks. The model reasons privately, then produces its visible answer:

markdown
<|im_start|>user
Your prompt here<|im_end|>
<|im_start|>assistant
<think>
… the model's private reasoning …
</think>

… the model's answer …<|im_end|>

Apply it automatically via tokenizer.apply_chat_template(messages, add_generation_prompt=True), or --jinja in llama.cpp.

Available formats

This repo — full-precision weights:

Table with columns: Format, ~Size, Best for
Format	~Size	Best for
BF16 safetensors	52 GB	transformers · vLLM · SGLang

GGUF quants → migtissera/Tess-4-27B-GGUF

Table with columns: File, Format, ~Size, Best for
File	Format	~Size	Best for
`Tess-4-27B-Q4_K_M.gguf`	Q4_K_M	16.5 GB	smallest — great quality/size · most popular
`Tess-4-27B-Q6_K.gguf`	Q6_K	22 GB	near-lossless
`Tess-4-27B-Q8_0.gguf`	Q8_0	28 GB	effectively lossless

Faster inference

⚡ Tess-4-27B-EAGLE3 — a speculative-decoding draft trained on Tess-4's own outputs: 1.76× average decode speedup, up to 2.4× on reasoning (measured on H100; lossless — outputs are identical). SGLang: --speculative-algorithm EAGLE3 --speculative-draft-model-path migtissera/Tess-4-27B-EAGLE3; vLLM: --speculative-config '{"method":"eagle3","model":"migtissera/Tess-4-27B-EAGLE3","num_speculative_tokens":4}'.
🧮 Tess-4-27B-NVFP4 — 4-bit NVFP4 (19 GB, −63%), Blackwell-native W4A4, calibrated on Tess-4's own generations. Quantization and speculative decoding stack.

Quickstart

llama.cpp / LM Studio (GGUF)

Grab the quant(s) from migtissera/Tess-4-27B-GGUF:

bash
hf download migtissera/Tess-4-27B-GGUF \
  Tess-4-27B-Q4_K_M.gguf mmproj-Tess-4-27B-F16.gguf \
  --local-dir ./tess-4-27b

bash
# text
llama-cli -m Tess-4-27B-Q4_K_M.gguf --jinja -p "Refactor this function and explain your reasoning."

# with images (multimodal)
llama-mtmd-cli -m Tess-4-27B-Q4_K_M.gguf \
  --mmproj mmproj-Tess-4-27B-F16.gguf \
  --image photo.png -p "What's in this image?"

transformers

python
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id = "migtissera/Tess-4-27B"
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": "Explain the tradeoffs of LoRA vs full fine-tuning."}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(inputs, max_new_tokens=1024)
print(processor.decode(out[0], skip_special_tokens=True))

(Requires a recent transformers with Qwen3.5/3.6 support.)

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.

Credits

License

Released under the Apache License 2.0, inherited from the base model. See LICENSE.

Citation

bibtex
@misc{tissera2026tess4,
  title        = {Tess-4-27B},
  author       = {Migel Tissera},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/migtissera/Tess-4-27B}},
  note         = {Built on Qwen/Qwen3.6-27B}
}

Tess-4-27B — part of the Tess series by Migel Tissera. Evaluations forthcoming.

Tess-4-27B-abliterated

README

Abliteration parameters

Performance

Tess-4-27B

Community Performed Benchmarks

References

Why Tess-4 is different

The reasoning traces

Prompt format & thinking

Available formats

Faster inference

Quickstart

llama.cpp / LM Studio (GGUF)

transformers

What it's good at

Credits

License

Citation

Explore FriendliAI today

README

Abliteration parameters

Performance

Tess-4-27B

Community Performed Benchmarks

References

Why Tess-4 is different

The reasoning traces

Prompt format & thinking

Available formats

Faster inference

Quickstart

llama.cpp / LM Studio (GGUF)

transformers

What it's good at

Credits

License

Citation