GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
wrldsuksgo2mars
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
wrldsuksgo2mars
Model Tree
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
Input Modalities
Output Modalities
Supported Functionality
The expert quantization used 1,426 prompts / 1,081,027 tokens spanning code review and rewriting, English and Chinese general text, multilingual text, math and reasoning, and structured tool calls. Quantization was propagated layer by layer: each layer's final K2/K3 mixture generated the activations for the next layer. The dSpark blocks used a deterministic 327,680-anchor sample of the jointly issued five-proposal target frontier.
The calibration corpus was text-only. The visual modules and visual routing biases are preserved exactly, but multimodal quality for this quantized expert set has not yet been measured.
DeepSeek-V4-Flash-Vision-Exp is DeepSeek's experimental multimodal V4 Flash
model. It combines the V4 Flash language architecture and built-in dSpark draft
path with a vision encoder and aligner. See the
DeepSeek-V4-Flash-Vision-Exp model card
for architecture details, prompt encoding, evaluation results, and license.
The model is provided under the MIT license from the source repository.