GLM 5.3 Flash Heretic

A community Heretic adaptation of GLM-5.3-Flash, available in original mixed precision and a compact native NVFP4 edition.

Based on Z.ai GLM-5.3-Flash, this release retains the complete expert network, native hybrid attention, multimodal components and MTP. Both editions contain the same targeted Heretic adaptation.

Check RELEASE_STATUS.json before downloading. A release is complete only when complete is true. Download one entire variant, not a single weight component.

Model highlights

• Reasoning, coding and tools: retains the base tokenizer, chat template and thinking controls.
• Native hybrid architecture: KDA linear attention, DSA sparse attention and manifold-constrained hyper-connections are retained.
• Complete MoE: 45 main layers, including 42 MoE layers with 288 routed experts each; all 12,096 main routed experts remain present, with eight selected per token.
• Multimodal and MTP assets: vision and multi-token prediction weights are retained without Heretic edits or additional quantization.
• Long-context architecture: the source configuration allows 1,048,576 positions. This is a configuration limit, not a new filled-context or quality result for this derivative.

Available formats

Original precision is not all-BF16. The actual source storage combines FP8 E4M3, BF16 and FP32. No full-BF16 conversion or additional rounding was performed for publication.

The NVFP4 weight data is about 40.6% smaller than the original-precision weights. Sizes above use decimal GB and exclude small indices/assets; disk size is not a runtime RAM/VRAM requirement. Exact file sizes and checksums are listed in FILES.json.

Heretic adaptation

The adaptation applies 21 targeted, norm-bounded residual-writer edits using GLM-specific captured activations. The selected working version uses an 8% per-selected-tensor relative edit cap. This describes the permitted weight displacement, not a percentage of quality loss, refusal reduction or modified parameters. It is not a universal ablation-strength standard.

The architecture and expert inventory are preserved. Vision and MTP are outside the edit. The NVFP4 edition retains all 21 modified matrices byte-for-byte in its protected native tensor pack.

Mixed NVFP4 recipe

The format is glm53-native-nvfp4-pack/2, codec glm53-nvfp4-g16-e4m3-mul-f32-v1. The complete NVFP4 edition requires both experts.bin and protected.bin, native-manifest.json, and assets/. Neither data file is a standalone model.

Compatibility: this is a custom native storage format, not GGUF and not a drop-in Transformers/ModelOpt NVFP4 checkpoint. A matching Strata loader/backend still requires implementation and qualification. No ready-to-run Strata engine or throughput result is claimed or included here. “NVFP4” describes stored weights, not guaranteed native FP4 Tensor Core execution on arbitrary hardware.

Validation and limitations

The original-precision checkpoint passed an independent 45-layer standalone check, matching candidate last-token logits bitwise and candidate NLL exactly under the declared FP8-weight-decode/BF16 reference arithmetic. This is a checkpoint-consistency check, not broad capability acceptance.

For the NVFP4 pack, an independent reader verified full data-file SHA256 hashes, asset completeness and tensor extents. All routed experts and protected matrices were accounted for. NVFP4 NLL/perplexity, generated-answer quality, vision/MTP execution and Strata runtime tests have not been performed.

No broad benchmark, unchanged-quality guarantee, refusal-rate score or measured performance is claimed for this GLM derivative. Heretic does not guarantee refusal-free, correct or hallucination-free answers. See EVIDENCE.json for the exact validation scope.

Download and use

from huggingface_hub import snapshot_download

# Choose exactly one complete edition:
snapshot_download(
    repo_id="XopMC513/GLM-5.3-Flash-Heretic",
    allow_patterns=["HereticOriginalPrecision/**", "LICENSE", "FILES.json", "RELEASE_STATUS.json"],
    local_dir="GLM-5.3-Flash-Heretic",
)
# For the custom native edition, replace the variant pattern with "NVFP4/**".

Keep each edition's files together and do not mix weight shards. Use a runtime that supports Glm5NextForConditionalGeneration and the edition's storage format. Preserve the bundled chat template and processor assets. The upstream template supports reasoning_effort values low, high, and max; upstream recommends clear_thinking=true for chat. MTP drafts must be verified by the target model.

Credits and license

Base model and original assets: Z.ai / zai-org, pinned revision 46a050bec13002ad29ee795c16e61cf4fa3c4a9c. Heretic adaptation and native quantized package: XopMC513 / Strata. This is an independent community derivative, not an official Z.ai release.

The original MIT license is retained.