GLM 5.3 Flash FP4 Local for Flare
This repository publishes a signed Flare runtime package for GLM 5.3 Flash, prepared from NVIDIA's FP4 checkpoint for two-GPU tensor-parallel serving on DGX Spark. Flare is a native Rust inference runtime. The package revision contains prepared weights, compiled kernels, captured execution graphs and signed package metadata. Transformers, vLLM, SGLang and llama.cpp cannot load this format.
The model handles text and images, accepts up to 1,048,576 tokens per
request, and serves an OpenAI-compatible API under the model name
GLM-5.3-Flash. Image requests have not been exercised on DGX Spark.
About Flare
Flare is NVIDIA's open-source inference runtime for local DGX and GPU systems. It is written in Rust and serves prepared, signed packages directly on the GPU using precompiled kernels and captured execution graphs, with no Python and no runtime compilation. It supports tensor-parallel serving across multiple hosts through Orbit, verifies publisher signatures before loading, and exposes an OpenAI-compatible API.
The weights
The package is built from
nvidia/GLM-5.3-Flash-FP4-Local,
NVIDIA's quantization-aware-distilled FP4 release of
zai-org/GLM-5.3-Flash, a
mixture-of-experts model with hybrid sparse and linear attention and
manifold-constrained hyper-connections: 320B total and 18B active parameters,
a 1,048,576-token context, and one MTP draft layer. Flare keeps the
checkpoint's NVFP4 routed experts with losslessly compressed scales, MXFP8
KDA projections and vocabulary head, and the replicated vision tower.
The FP4 checkpoint is governed by the
NVIDIA Open Model Agreement
(LICENSE); the GLM 5.3 Flash base model is under the MIT license
(LICENSES/MIT-GLM-5.3-Flash.txt). See NOTICE.
Pick a platform
| Revision | Hardware | Package | Platform setup |
|---|---|---|---|
gb10 |
2× DGX Spark (GB10, SM121) across two hosts | ~168 GB | gb10 setup |
The main revision holds this model card, the licenses and the publisher
profile. Download the gb10 revision; its README covers the Flare build it
needs, launch, storage and the measured capacity.
Download and verify
Download with flare download or the Hugging Face hf CLI, always naming the
revision. Flare verifies the package signature against publisher.json, then
asks you to accept the publisher and the model license. The signature covers
the metadata primary and every companion, so a missing or renamed payload fails
verification.
Comprehensive model-accuracy evaluation has not been run.