0xzknw/Ling-3.0-tiny-Heretic-NX-PRIME-GGUF

Ling 3.0 Tiny — Heretic NX PRIME Q8_0

Heretic NX PRIME is a direct-Q8 behavioral edit of `inclusionAI/Ling-3.0-tiny`, a 7.9B-total / 1.3B-active hybrid KDA–MLA MoE model. It is designed to reduce false refusals while explicitly measuring drift from the unedited Q8 deployment artifact.

The release contains the evaluated Q8_0. PRIME is the project's internal evidence and capability-preservation protocol, not a third-party certification.

File

Evaluation

The original and edited GGUFs were loaded from disk with the same llama.cpp Metal build. Refusal evaluation used the native pre-tokenized /completion endpoint, greedy sampling, 96 generated tokens, four runtime slots and the official chat template with enable_thinking=False.

The release satisfies both predeclared targets: at most 6 refusal markers and mean first-token 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 candidate selection, 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. Thinking was disabled for both.

The paired bootstrap 95% interval for Heretic minus original is [-2.93, -0.12] points. There were 655 questions both got right, 164 both got wrong, 24 original-only successes and 11 Heretic-only successes. This passes the predeclared 3-point non-inferiority margin and the symmetric ±3-point equivalence gate on this slice. It does not demonstrate an accuracy increase.

Method

The edit uses benign-protected, per-site rank-8 directions measured from the official BF16 model as research instrumentation, then merges the selected projectors directly into the source Q80 GGUF at `beta=2.70`. The released artifact itself remains Q80 throughout the edit and evaluation path.

Only 16 dense shared operators are edited: attention outputs in layers 10, 11, 12, 13, 14, 15, 22 and 23, plus shared-expert FFN down projections in layers 11, 12, 13, 14, 15, 16, 17 and 21. No routed MoE expert-bank tensor is edited.

The source GGUF is `bartowski/Ling-3.0-tiny-GGUF` Ling-3.0-tiny-Q8_0.gguf, SHA-256 628fde593836ab3fd291e93802c8699c7e4eb3e32ba2562e05ba7b906a93be5e. The official BF16 research checkpoint was pinned to revision b61f4338de3e68ffc9c0bc1ed5e902981a4a929e; the source GGUF repository was pinned to ea072726af0d2e8ba325b2f90fc0efa762105a91.

All final metrics were computed with llama.cpp commit 18443257a30c884d5332abb8e7dc43c7ffe42fda. The exact plan, merge report, rank-8 factors and evaluation reports are included under evaluations/. Raw-logit collectors and reproducible experiment scripts are available in `0xZKnw/heretic-nx`.

LM Studio / llama.cpp

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

llama-server \
  -m Ling-3.0-tiny-Heretic-NX-PRIME-Q8_0.gguf \
  -ngl 99 -c 4096 --jinja

Use a recent runtime with bailingmoe3 support. Ling 3.0 Tiny supports both thinking and non-thinking modes; applications should pass the desired enable_thinking chat-template option explicitly.

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 shows a small aggregate accuracy decrease. It should not be read as a universal quality guarantee. The release is not externally certified or claimed equivalent to the original model on every task.

The base model is licensed under the MIT License.