763B DeepSeek V4.1 Flash 2-bit runs at ~50 tok/s on a single 32 GB GPU via FLE
Ryyn_- · reddit · 2026-11-07
A developer demoed FLE Link + FLE engine serving a 2-bit quantized DeepSeek V4.1 Flash (Rin) — 763B parameters compressed onto a single 32 GB GPU at roughly 50 tok/s.
FLE Link exposes OpenAI- and Anthropic-compatible endpoints; flelink connect claude prints ready-to-use configs for Claude Code, Codex, OpenCode, and pi. The engine offers a /fle/status endpoint (cache, draft acceptance, live tok/s, warnings) and reports startup failure reasons like fitbudgettoolarge or ramtoosmall. MIT-licensed and installable with one command. The author discloses they make FLE.
More from coding & agent
- When code is nearly free: the real bottlenecks in scaling AI-built software — rseroter · 2026-11-07
- OpenClaw rebuilt model discovery so new models ship without waiting for a release — heyneighbor · 2026-11-07
- Jiayuan shares his agent merge workflow prompt with confidence-scored triage — jiayuan_jy · 2026-11-07
- OpenAI Agents API October drop: auto review, scheduled automations, mid-session tool swaps — OpenAIDevs · 2026-11-07
- A real unlock: make coding agents put code in the right folders — structure is documentation — MikkoH · 2026-11-07
- Let Codex clean up after itself: weekly reminder to clear Rust build cache on Mac — gethackteam · 2026-11-07