SYLIN

DeepSeek

Maker · 11 entries

Everything from DeepSeek

NameWhat to knowPrice or memoryRanks
DeepSeek FlashModelstrong Best-value open-weights model: near-frontier work for a few dollars, though a step behind.~510 GB#2 web development (open weights)
DeepSeek app-feature APIApp or serviceDeepSeek app features exposed via API: search, files, OCR, ASR, TTS, with separate tool fees and an open-source frontend
DeepSeek HarnessAgentpromising Free, MIT-licensed coding harness that ties for first on a leaderboard, with one scoring flaw.
DeepInferRuntimeunproven DeepSeek's serving engine is open, but its production economics can't be reproduced.
DeepSeek ProModelUntil V4.1 Pro ships, all V4 Pro requests are routed to V4.1 Flash and billed at Flash prices~893 GB
CSA2TechniqueFull, reindex and reuse layer modes share KV cache and indices across layers
DSparkTechniqueMulti-token-per-step decoding component included in V4.1 Flash
EngramTechniqueSeparate static-fact memory module held in server RAM; 168B params per video (196B per HF page)
Single-pass mHCTechniqueFuses operations to cut intermediate GPU memory traffic
SWA bounded replayTechniqueDiscards local sliding-window cache after each reply and recomputes the last 128 tokens on the next turn
DeepSeek Flash VisionModelExperimental vision version of V4 Flash, open-sourced on Hugging Face after earlier API release; 305B parameters, about 168 GB~168 GB