GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
appmana
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
appmana
Model Tree
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
Input Modalities
Output Modalities
Supported Functionality
The IMMA checkpoint scored 100/100 on a fixed suite of controlled image fixtures; the matched BF16 vision reference scored 99/100. This is a regression check, not a general vision benchmark. Eight simultaneous requests with eight distinct images each completed. A stretched-image stress variant completed with identical counting errors in four of eight requests in both variants.
The IMMA runtime completed a 990,160-token image prompt. Its repeated prompt reused 989,952 tokens (99.979%) through LMCache. A separate image replay after restarting all model/cache pods recovered 5,120 tokens from Redis on every rank with initially empty local caches. Redis capacity was bounded to 12 GiB; full long-context persistence in Redis is not claimed.
Minimum sampled free GPU memory across startup and tested IMMA workloads was 1.161 GiB per GPU or greater. Configured limits were 1,000,000 context tokens, eight concurrent sequences, and eight images per request. Production SM86 PTX verified signed INT8 tensor-core MMA in vision linear and attention kernels.
Validated weight revisions: IMMA 65df4c51a2e724340c7963ccddd1f5199a2575e4;
BF16 reference bee7f4e9f445e49067b35d88e4ef3ef3ca9f1a56.
Runtime: AppMana vLLM b281385271db074c1448a98dfc3ad61a18db34a0,
LMCache d55f82dd97d919876e396598260cfd315503e854.