GLM-5.3-Flash-Maya-M-Derisked-IQ2_S-GGUF

This is the Maya-M IQ2_S release of GLM-5.3-Flash with a targeted Directional Weight Modification (DWM) applied directly to the GGUF weights.

The DWM was performed by Blackfrost_AI using a proprietary method designed to reduce blanket refusal behavior while preserving the model's original GGUF identity and packaging.

This model is not a requant. It retains the source IQ2_S mixed-storage quantization, tensor names, tensor count, tensor shapes, storage types, shard layout, tokenizer, chat template, NextN draft layer, and exact aggregate byte size. DWM intentionally changes the numerical values of selected target tensors; “preserved identity” does not mean the edited weights are bit-identical to the source.

Model summary

What was changed

Blackfrost_AI applied DWM to measured residual-write surfaces to reduce over-refusal in authorized specialist workflows. The modification was performed inside the source GGUF representation. Selected matrices were transformed in working precision and written back once to their original storage types.

This release is:

• a direct derivative of the named Maya-M IQ2_S source;
• not a whole-model requant or quantization conversion;
• not an SFT, DPO, distillation, expert-pruning, or router-surgery release;
• structurally compatible with the same GLM-5-next-capable GGUF runtime class as the source.

The proprietary capture data, direction vectors, prompt pairs, transformation parameters, implementation, internal logs, and evaluation bundle are not part of this repository.

Structural validation

The release was checked against the immutable source after the edit:

• all three source shards retained their pre-build hashes;
• tensor names, count, shapes, dtypes/storage types, ordering, and split layout matched the source;
• all intended target tensors changed and non-target payloads remained byte-exact;
• all 29 NextN/MTP tensors were retained byte-for-byte and were excluded from the DWM edit surface;
• the output shard hashes match SHA256SUMS;
• the model loaded successfully in a GLM-5-next-capable llama.cpp runtime;
• an OpenAI-compatible runtime smoke test passed at 131,072 context with reasoning separated from final content.

Evaluation status and limitations

This is an experimental weight-edited release, not a safety certification. Structural and runtime checks passed. A clean-versus-edited, frozen held-out, full-vocabulary distribution-drift evaluation has not been completed, and no broad benchmark or production-readiness claim is made here.

Reducing over-refusal can change model behavior in ways that are task- and prompt-dependent. Deployers should run their own capability, safety, domain, tool-use, and long-context evaluations and provide application-level access controls, monitoring, and policy prompts appropriate to their use case.

Files and integrity

After downloading all files, verify them with:

sha256sum -c SHA256SUMS

Deployment

Point llama-server at the first shard; compatible runtimes discover the other two automatically. Thinking is enabled by the embedded template, so allow a generous output budget when reasoning is requested.

Known-good serve shape:

llama-server \
  --model "$MODEL_DIR/GLM-5.3-Flash-Maya-M-Derisked-IQ2_S-00001-of-00003.gguf" \
  --alias GLM-5.3-Flash-Maya-M-Derisked-IQ2_S \
  --ctx-size 131072 \
  --n-gpu-layers 99 \
  --parallel 1 \
  --flash-attn off \
  --fit off \
  --jinja \
  --chat-template-kwargs '{"reasoning_effort":"max"}' \
  --temp 1.0 \
  --top-p 0.95 \
  --host 0.0.0.0 \
  --port 8000

OpenAI-compatible routes are GET /v1/models and POST /v1/chat/completions.

Lineage and credit

• Base model: zai-org/GLM-5.3-Flash
• Maya-M IQ2_S GGUF: peasantsmith/GLM-5.3-Flash-Maya-GGUF
• GGUF-native DWM: Blackfrost_AI

The original GLM-5.3-Flash model and the Maya GGUF source are MIT licensed. The included LICENSE is the upstream Z.ai MIT license and applies to this derivative.