No coding agent vendor has shipped DeepSeek's open-sourced DeepInfer yet
pingzawa162 · reddit · 2026-10-26
DeepSeek recently open-sourced DeepInfer (its in-house inference engine) and deepseek-recipe (which keeps self-hosted output identical to the official API), but no coding agent vendor has publicly adopted either yet.
- DeepInfer serves V4.1 Flash at 155T tokens/day per cluster, 400ms first-token latency, 300 tok/s for coding
- deepseek-recipe ensures self-hosted deployments produce the same token IDs and format as the official API
- V4.1 Flash 1017 offers 1M context at $0.30/M input, $1.20/M output, $0.26 per task
- Hardware is locked to SM100 (Blackwell) and Ascend 950; only V4.1-Flash is supported
- SGLang is already integrating deepseek-recipe, but product-level adoption remains absent
Bottom line: for now, the official API is the only practical option for most users.
More from coding & agent
- RAG Reranker Showdown: Small Cross-Encoder Pushed Correct Chunks Out of Top 5 — Affectionate_Slip654 · 2026-10-26
- 10 open-source GitHub repos that give your AI agent a chat UI, by stars — victor_explore · 2026-10-26
- SirenoRELAY: a policy-boundary MCP that lets AI agents read only allowed mail and send only to allowed recipients — ismaelfaro · 2026-10-26
- Athena Crisis migrates off Relay: 1k fewer lines, 3.7% smaller bundles, ~400ms faster page loads — cnakazawa · 2026-10-26
- Gartner chart: 12 enterprise tech roles collapsing into Product Engineers and AI platform teams — luisdans · 2026-10-26
- YOURLS-MCP lets Claude Desktop manage your self-hosted YOURLS shortener — modelcontextprotocol · 2026-10-26