DeepSeek 552B model runs on a single AMD MI325X: 4 agents at 100 tok/s
realmrfakename · x · 2026-10-25
lxe demonstrates running DeepSeek-V4.1-Flash-1017 (552B params + 196B Engram table) on a single 256GB AMD Instinct MI325X using his dsx-amd engine (Rust + HIP).
- 100 decode tok/s per session across four concurrent coding agents, full expert routing with no pruning
- Native MXFP4 experts mixed with fitted 3-bit experts to fit VRAM; Engram table on NVMe + host page cache
- 90.6% token agreement with FP8 on the Lucebox eval (88.3% in serving profile)
- OpenAI/Anthropic-compatible API with streaming and tool calls
More from coding & agent
- Base44 adds Entri integration: connect custom domains to your apps in a couple of clicks — tristanbob · 2026-10-25
- Grok Bot burns a week of Cursor Ultra quota in one day for a $200/mo user — alexcovo_eth · 2026-10-25
- Agent instruction files are executable code that skips code review, warns Cairnlock dev — augustkevinsanjaya · 2026-10-25
- Redditor builds unattended ticket-to-code pipeline with Linear and Claude Code — Awkward_Rent4749 · 2026-10-25
- dotey: Grokbot beats Claude Code and Codex for on-the-go mobile agent use — dotey · 2026-10-25
- 10 open-source GitHub repos that give your AI agent the whole internet — victor_explore · 2026-10-25