fiore-lab/GLM-5.3-EXL3-Base

English | 日本語 | 简体中文

GLM-5.3, EXL3: the weights a start reads beside the expert banks

Everything of zai-org/GLM-5.3 (755B, glm_moe_dsa) but the routed experts of layers 3 to 77, by Fiore-lab, for the recipe that serves it on NVIDIA DGX Spark systems: https://github.com/fiore-labo/ExpertAtlas-GLM5.3 (release r3.3-102 or later). That recipe's engine is the only one these files are laid out for.

Every start needs this repository and the 3-bit bank, fiore-lab/GLM-5.3-EXL3-Experts-3bit: 299.1 GB of disk on each node together, on four nodes and on eight. The engine reads each routed expert from the banks, at the width chosen expert by expert. The full checkpoints, fiore-lab/GLM-5.3-EXL3-8x-DGX-Sparks and fiore-lab/GLM-5.3-EXL3-4x-DGX-Sparks, hold the routed experts as well; a start does not need them, and either works in place of this repository.

What is in it

File	Contents
nonexperts-5bit.safetensors	the other linears of layers 0 to 77 (attention projections, shared experts, the dense MLP of layers 0 to 2, the indexer's wq_b): EXL3, 5 bits, codebook mul1
unquantized-bf16.safetensors	the tensors kept as the BF16 checkpoint stores them: embeddings, lm_head, norms, router, kv_b_proj, the indexer's wk / weights_proj / k_norm, the final norm (byte for byte the official ones)
mtp-layer-78-*.safetensors (3)	the MTP layer (layer 78), the recipe's default drafter: routed experts 4 bits mcg, linears 5 bits mul1, eh_proj and the other unquantized tensors BF16
generation_config.json, tokenizer.json, tokenizer_config.json, chat_template.jinja	the official files
config.json	the official file with a quantization_config; its routed_experts says the banks hold them
model.safetensors.index.json	the tensor index without the routed experts of layers 3 to 77: 6,014 names
SHA256SUMS	the checksums of every file but this card
LICENSE	GLM-5.3 License

13 files without this card, 21.216 GB (19.76 GiB) in all. The five .safetensors files, the small official files and LICENSE are byte for byte those of the two full checkpoints; config.json, the index and SHA256SUMS are its own. It holds no draft model but the MTP layer, and no routed expert of layers 3 to 77.

To check a download, run this in its folder:

sha256sum -c SHA256SUMS

How it was made

From the four-node checkpoint, with the recipe's assemble.py base: the files above as they are, the index without the routed experts, and config.json saying where they are. Nothing was quantized again.

Making it yourself

Every file here can be made from zai-org/GLM-5.3-BF16 (revision 9d2398f) with the recipe's tools, without downloading this repository: glm53/weights. assemble.py base makes this layout from the eight-node checkpoint you assemble there.

License

The GLM-5.3 License (LICENSE), Copyright (c) 2026 Z.AI, as for the base model: the copyright notice and the permission notice must be included in all copies or substantial portions. Its clause 2 reads: "If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." Check whether it applies to you. The recipe's code is Apache-2.0 and licensed separately.