LFM2.5 8B-A1B — Heretic NX PRIME GGUF

Heretic NX PRIME is a direct-GGUF behavioral edit of LiquidAI/LFM2.5-8B-A1B. It is designed to reduce false refusals while explicitly measuring drift from the unedited deployment artifact.

The release is the evaluated UD-Q8_K_XL. PRIME is the project's internal evidence and capability-preservation protocol, not a third-party certification.

Files

Evaluation

Both final files were loaded from disk with llama.cpp and evaluated over all 104 rows. Decoding used the native pre-tokenized /completion endpoint, greedy sampling, 96 generated tokens, the closed-thinking chat template and four runtime slots.

The release satisfies both targets: at most 6 refusal markers and mean KL at most 0.05.

Lexical markers are a refusal proxy, not semantic task success or a universal quality score. All 104 harmful rows participated in development and selection, so this suite is not an untouched holdout.

Paired capability check

After the Q8 candidate was frozen, the original and Heretic Q8 artifacts were evaluated on the same deterministic 854-question slice from ARC-Challenge, HellaSwag and MMLU. Each answer was the greedy first-token argmax restricted to A/B/C/D; prompts, tokenizer, runtime and precision were identical between arms.

The paired bootstrap 95% interval for Heretic minus original is [-1.29, +1.29] points. There were 454 questions both got right, 368 both got wrong, 16 original-only successes and 16 Heretic-only successes. This passes the predeclared 3-point non-inferiority margin and the symmetric ±3-point equivalence gate. It does not demonstrate an aggregate accuracy increase; it supports capability preservation on this narrow multiple-choice slice.

Method

The selected edit is a benign-penalized distillation of an eight-site PRIME teacher. It fits conditional rank-one right factors from 1,024 harmless states and 2,627 response-trajectory states, then merges them directly into eight Q8_0 operator-output tensors at lambda=100 and beta=2.25.

No MoE expert-bank tensor is edited. The eight sites are two attention outputs and six short-convolution outputs in layers 12, 14, 16, 17, 19, 21, 22 and 23. Every plan, factor artifact and output is SHA-256 bound. The direct-Q8 backend, tests and reproducible experiment scripts are available in 0xZKnw/heretic-nx.

LM Studio / llama.cpp

Download one GGUF and load it normally. For llama.cpp:

llama-server \
  -m LFM2.5-8B-A1B-Heretic-NX-PRIME-UD-Q8_K_XL.gguf \
  -ngl 99 -c 4096 --jinja

A recent runtime with lfm2moe support is required.

Limitations and responsibility

This edit intentionally weakens refusal behavior. It can increase compliance with unsafe, illegal, incorrect or otherwise harmful requests. It does not add factuality, judgment, sandboxing or application-level safety. Run untrusted generations in an appropriate sandbox.

The paired capability check is narrow and should not be read as a universal quality guarantee. The release is not claimed to be a universal winner, externally certified or equivalent to the original model on every task.

Use is subject to the included LFM Open License v1.0.