GLM-5.3-Flash-MXFP4-mtp

Model architecture

A 320B-class mixture-of-experts multimodal model with about 18B active parameters per token. This conversion retains the text backbone, vision encoder, and one multi-token prediction (MTP) layer in a custom packed binary format.

Component	Structure
Text backbone	45 layers; hidden size 4,096; the first three layers use dense feed-forward networks of size 12,288
Attention layout	34 KDA linear-attention layers and 11 sparse-attention layers, with sparse attention every fourth layer
Sparse attention	64 heads; query/key and value head dimensions 256; compressed KV dimension 512; no rotary query/key dimensions
IndexPool	Groups four index keys into one pool; configured selection budget of 2,048 tokens, plus the incomplete tail
Routed experts	42 main-model layers, each with 288 experts; eight selected per token; expert intermediate size 2,048; one shared expert per layer
Residual connections	Four-stream manifold-constrained hyper-connections (mHC), with 20 Sinkhorn iterations
Vocabulary	154,880 tokens
Configured context limit	1,048,576 tokens; this conversion has not been validated at that length
Vision encoder	24 layers; hidden size 1,024; 16 attention heads; 14 × 14 spatial patches
Vision-to-text connection	Vision features are projected to the text hidden size of 4,096
MTP	One additional prediction layer; its weights and routed experts are included

Weight precision

This checkpoint uses mixed precision rather than uniform 4-bit weights:

• All 12,384 routed experts, including the MTP layer's experts, are converted from source FP8 to MXFP4 E2M1, with 32 values per group and one E8M0 scale byte per group.
• Each expert contains gate, down, and up projections. Packed values occupy four bits each; scales add an effective 0.25 bits per weight, for 4.25 bits per routed-expert weight before file metadata.
• Source FP8 weights use FP32 scales for 128 × 128 blocks. Conversion applies these scales in FP32, then quantizes directly to MXFP4 without an intermediate BF16 rounding step.
• The group exponent is floor(log2(max_abs)) - 2, clamped to the supported scale range. Values use nearest-even E2M1 rounding. This is a lossy conversion.
• Non-routed weights, including shared experts, dense feed-forward networks, attention projections, embeddings, normalization, vision weights, and MTP non-routed weights, retain their source FP8, BF16, or FP32 representations as applicable.

The payload contains 182,191,831,203 bytes (about 182.19 GB / 169.68 GiB) across 46 binary files and eight source companion files. Repository documentation and conversion metadata are additional.

Format and validation

This is a custom binary artifact, not an MLX, GGUF, or directly loadable Transformers checkpoint. A compatible loader and inference implementation are required. The original files under checkpoint/ describe the source model, not a ready-to-load set of safetensors in this repository. In particular, its original shard index refers to source shards that are not included here.

Path	Contents
common.bin	Main-model non-routed weights
experts/layer_03.bin through experts/layer_44.bin	Main-model routed experts
mtp/common.bin, mtp/experts/layer_45.bin	MTP weights
vision/common.bin	Vision encoder weights
checkpoint/	Original configuration, tokenizer, chat template, processor configuration, shard index, and license
repack-plan.json	Tensor shapes, source ranges, destination offsets, and conversion operations
conversion.json	Quantization settings, payload sizes, and SHA-256 digests
source-info.json	Pinned source file metadata and digests

Each routed expert occupies 13,369,344 bytes. Expert IDs are stored consecutively within each layer file. Within each expert, gate, down, and up projections each store packed row-major weights followed by their group scales; the low nibble represents the earlier column. Exact offsets are recorded in repack-plan.json.

All 62 source weight shards passed SHA-256 verification before conversion. All 54 converted payload and companion files passed read-back SHA-256 verification. Nine sampled projections from early, middle, and late layers also matched separately checked conversion results byte-for-byte.

Full-model inference, task quality, generation speed, long-context behavior, image inference, and MTP decoding have not been validated for this artifact. Included weights do not imply working runtime support. qualifiedForInference remains false in the conversion metadata. No benchmark scores from the source model are claimed for this conversion.

Source and license

Source: zai-org/GLM-5.3-Flash.

Source revision: 46a050bec13002ad29ee795c16e61cf4fa3c4a9c.

Original model by Z.ai. This conversion retains the upstream MIT license. The original license is also preserved under checkpoint/LICENSE.