Modelpublic

Blackfrost-AI/GLM-5.3-DERISKED-Int4-Int8Mix

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

4likes1.1kdownloads

Model Card

GLM-5.3-DERISKED-Int4-Int8Mix

Full GLM-5.3 MoE · Tech2wild Int4-Int8Mix quant · Blackfrost DWM

Built by [Blackfrost](https://x.com/Blackfrost_AI) · Las Vegas, Nevada

</div>

Public and ungated. Same derisked weights Blackfrost applied on top of Tony / Tech2wild's Int4-Int8Mix quant.

Credit — Tech2wild

This checkpoint starts from [Tech2wild/GLM-5.3-Int4-Int8Mix](https://huggingface.co/Tech2wild/GLM-5.3-Int4-Int8Mix) by Tony ([@Tech2Wild](https://huggingface.co/Tech2wild) / [tonyd2wild](https://github.com/tonyd2wild)). That is the Int4-Int8Mix compressed-tensors quant of full GLM-5.3 that actually fits 4× NVIDIA DGX Spark. The Hub base model for this repo is that Tech2wild artifact.

Recipe, serving notes, and Spark TP4 numbers for the stock quant live in Tony's writeup: tonyd2wild/GLM-5.3-Int4-Int8Mix-TP4-4x-DGX-Spark.

Original architecture / license parent: [zai-org/GLM-5.3](https://huggingface.co/zai-org/GLM-5.3) (Z.AI). Quant format: Neural Magic / Red Hat compressed-tensors. Layer map: QuantTrio-style Int4 experts / Int8 dense+attention, as reproduced by Tech2wild.

Blackfrost did not re-quantize these weights. We copied the Tech2wild tree and applied an independent direction-weight modification (DWM) pass. Stock Tech2wild shards were not overwritten.

What this is

Independent Int4-Int8Mix pack-quantized safetensors of GLM-5.3 (full ~743B MoE / ~40B active, not Flash) after a Blackfrost DWM pass.

The intended behavior is in the weights. Production DWM details are proprietary and are not disclosed beyond the locked recipe below.

This artifact has not been through a judged refusal suite. Do not copy NVFP4 / BF16 / Flash GGUF refusal percentages onto this checkpoint.

Specifications


Architecture	`GlmMoeDsaForCausalLM` (`glm_moe_dsa`)
Quant source	`Tech2wild/GLM-5.3-Int4-Int8Mix` · rev `206507bbb047d8223964a0414cd83230c59428f9`
Quant	`compressed-tensors` `pack-quantized` · int8 group-128 linears · int4 group-128 routed experts · layer 0 BF16 · MTP int8 channel (stock Tech2wild)
Shards	282 safetensors
Layers	78 main + MTP layer 78
Hidden size	6144
Experts	256 routed · top-8 / token · 1 shared
DWM	alpha 3.0 · skip-early 2 · one pass · frozen scales · no norm restore
Targets	76 `o_proj` + 1 dense `down_proj` + 19200 expert `down_proj` + 75 shared `down_proj`
Runtime	vLLM `compressed-tensors` / pack-quantized (multi-Spark). Not a GGUF.
Languages	English and Chinese
Access	Public, ungated

Serving

Same vLLM compressed-tensors path as the Tech2wild stock quant. GB10 / sm121 still needs the Spark kernel overlays Tony documents; on standard CUDA, --quantization compressed-tensors is the load flag.

bash

vllm serve Blackfrost-AI/GLM-5.3-DERISKED-Int4-Int8Mix \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --tensor-parallel-size 4 --enable-expert-parallel \
  --served-model-name glm-5.3-derisked \
  --max-model-len 131072

Disclaimer

This checkpoint has a deliberately altered refusal profile and is intended for research and local use. It is not a safety boundary. Outputs may be inaccurate, offensive, unsafe, or otherwise unsuitable. Operators are responsible for lawful use and application-level controls.

License

Inherits the GLM-5.3 license from Z.AI. Quantization and DWM change weights only, not the license terms. Preserve upstream copyright and license notices.

Contact

@Blackfrost_AI on X · Blackfrost · Las Vegas, Nevada