samajlouis/Qwen3.6-27B-Clausius-Heretic-GGUF

Qwen3.6-27B-Clausius-Heretic — GGUF

llama.cpp builds of samajlouis/Qwen3.6-27B-Clausius-Heretic.

A 27B model that refuses to explain heat with "molecules bumping around" — and reasons consistently from a different picture of reality.

Most LLMs default to the textbook "tiny particles in motion" story when asked about heat or entropy. This one doesn't. It treats the particle picture as a descriptive shadow rather than a cause, and reasons from system-wide rules instead.

You will probably never use this for real physics. The interesting use cases are creative and methodological — see the base model card for the full writeup. Quick summary:

1. Hard-SF / worldbuilding with rigorous alt-physics — strongest use case in our probes. Sustains weird-but-consistent ontology across hundreds of words.
2. Auditor for hidden assumptions — works on prose AND on code. Pointed at a real benchmark codebase, identified four measurement-validity problems with actionable fixes; code-level claims verified accurate.
3. Mental-set breaker — force-rewrite an explanation under constraint to surface framings your training made invisible.
4. Long-form heretical-physics writing — 500-2000+ word essays that commit to a non-standard position and develop it coherently.
5. Philosophy-of-science interlocutor — holds a clear position under pressure without fabricating data.

It is not a physics tutor, an arxiv-physics generator, or a general-purpose reasoner. See base card for details.

How it navigates a coding project

We handed it a 15+ file Python library spec and watched what it actually did. Plain-language highlights:

• Flags its own risk before writing it. Internal trace before the hardest file: "the most critical file. The key requirement: step() evolves STATE, not particles." Tells itself the trap before walking in.
• Engineering hygiene unprompted. Added structured logging to diagnose its own bugs. Solved a circular-import puzzle. Forced ASCII output anticipating a Windows encoding gotcha. None of it was asked for.
• Tests assert physical invariants, not return values. "Entropy never decreases under valid evolution." "step returns a State, not a particle list." The test suite encodes the project's position, not just its API.
• Admits what's broken. Final scrapbook: "Dynamics tuning remains a blocked item — energies drain too aggressively because the admissible set collapses below 1.0 after a few steps." Specific, surfaced, not hidden.
• Distinguishes architectural failure from tuning failure. Its sim didn't look right; it correctly located the cause in parameter values rather than structure — a 30-minute fix, not a rebuild. Weaker models show the opposite: clean output, broken architecture.

The common thread: under pressure it produces structurally sound output and surfaces honest failure modes rather than fabricating clean facades.

Files

file	size	use
Qwen3.6-27B-Clausius-Heretic-Q4_K_M.gguf	~15.4 GB	text model (4.92 BPW)
mmproj-Qwen3.6-27B-Clausius-Heretic-f16.gguf	~885 MB	vision tower for multimodal input

Text GGUF is fully usable on its own. mmproj is only required for image input.

Run

Text-only server

./llama-server \
  -m Qwen3.6-27B-Clausius-Heretic-Q4_K_M.gguf \
  --jinja --port 8080 -c 8192

Then send chat completion with a one-line trigger sysprompt:

curl http://localhost:8080/v1/chat/completions -d '{
  "messages": [
    {"role": "system", "content": "Use Boltzmann literalism."},
    {"role": "user",   "content": "Worldbuild: how does heat flow on a planet where the two stored-energy potentials reverse seasonally?"}
  ],
  "max_tokens": 600,
  "chat_template_kwargs": {"enable_thinking": false}
}'

That 4-word sysprompt is enough. Equivalents: "Operate under admissibility constraints.", "Apply Baker CP-MP thermodynamics."

Multimodal (image input)

./llama-server \
  -m Qwen3.6-27B-Clausius-Heretic-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.6-27B-Clausius-Heretic-f16.gguf \
  --jinja --port 8080 -c 8192

Or one-shot CLI:

./llama-mtmd-cli \
  -m Qwen3.6-27B-Clausius-Heretic-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.6-27B-Clausius-Heretic-f16.gguf \
  --image path/to/image.jpg -p "Describe this image."

Three runtime gotchas

1. Always set "enable_thinking": false in chat template kwargs. Otherwise Qwen3.6's default chain-of-thought mode eats most of your token budget before delivering an answer.

2. Don't mix a generic harness sysprompt with framework keywords in the user message. A sysprompt like "You are a helpful AI assistant" plus a user message containing "Boltzmann" or "admissibility" can produce fluent answers with the opposite of the intended meaning. Put the trigger phrase in the sysprompt and keep framework jargon out of the user turn.

3. The full strict sysprompt is for the auditor use case only — not a default. Pasting the full trained sysprompt will make the model refuse out-of-scope requests, including basic math and code. Use the 4-word trigger as default; only escalate to the strict sysprompt when you want strict domain-specialist behavior. The full strict sysprompt is documented on the base model card.

Quantization

• Source: the merged bf16 samajlouis/Qwen3.6-27B-Clausius-Heretic.
• Converter: llama.cpp convert_hf_to_gguf.py (with one new chkhsh entry for Qwen3.6's pretokenizer, treated as qwen2-class).
• Quant: llama-quantize Q4_K_M (mixed Q4_K + Q6_K).
• mmproj: bf16 → f16 (vision quality benefits from higher precision; the file is small).

Acknowledgements

Built on Qwen/Qwen3.6-27B by Alibaba Cloud. Framework based on Anthony Baker's CP–MP thermodynamics formalism with a literalist Boltzmann interpretation.