FYOS

Reports · films · releases

Everything we
have published.

Technical reports with raw data, architecture documents, films about the measurements, and every package, image and model that is publicly available.

ReportsFilmsReleasesTimeline

Technical reports and documents

Technical report FLE: Fast Local Experts An 89 GB MoE model at ~90 tok/s on one 32 GB GPU: quantization, expert cache, MTP, prefill, results against upstream and Strata, standard benchmarks, negative results and every paired bracket in the appendix. v1.0.2 · 2026-10-10 · doi:10.5281/zenodo.23272400 HTML ↗PDF ↗DOI ↗
Model card Qwen3.8-Flash-Next-FLE-GGUF Three hybrid quantizations with per-layer provenance, the MTP head and the eight films. FLE-IQ3_M · IQ3_XXS · IQ2_M · MTP
Model card Qwen3.8-Flash-Next-FLE-REAP448-GGUF FLE-IQ3_M with 64 of 512 experts per layer pruned – and a knowledge probe showing what KLD understates. REAP · 448 / 512
Architecture ARCHITEKTUR.md How shell, hub, agent and plugins work together and where the boundaries between them lie. fyos-os · docs
Contract HUB-VERTRAG.md The hub’s routes and formats: MCP, REST, apps, health. fyos-os · docs
Guide ERWEITERN.md Hub plugin, app, skills, agent behaviour, shell, catalogue, companion services, testing and shipping – with examples. fyos-os · docs
Coupling KOPPLUNG-OPENCLAW.md How the agent builds on OpenClaw without forking upstream. fyos-os · docs
Behaviour contract fyos-agent/vertrag What the agent may do, when it asks and when it speaks up – as prose and machine-readable. Betriebskonzept · Freigabe · Meldungen · Datenschutz
Cite

Max Bitzer (sirfyyn). FLE: Fast Local Experts — an 89 GB Mixture-of-Experts model at ~90 tokens/s on one 32 GB GPU. Technical report, 2026. doi:10.5281/zenodo.23272400

Making Rhea Fast

Eight short films (English, 33 to 63 seconds, 7 minutes in total) show how FLE was measured on one machine and what did and did not work. Every number comes from the measurement database behind the report (6,130 measurements); where a film and the report differ, the report is the reference.

0:33 1/8 The Race 0:34 2/8 The Bracket 1:03 3/8 The Staircase 0:59 4/8 The Expert Cache 1:00 5/8 Draft and Verify 0:56 6/8 What Did Not Work 0:58 7/8 The Rebench 0:58 8/8 Why We Do This

All eight as one film (6:59) ↗ Playlist on YouTube ↗ English, subtitles burned in. Links go to YouTube.

Releases and artefacts

Checked on 2026-10-21. The distribution is Apache-2.0, FastLocalExperts is MIT; the models follow the base model’s license.

PyPI

fyos-hubMCP hub with bundled plugins0.1.3
fyos-installerinstaller: web, text UI, command line0.1.3
fyos-plugin-vorlagetemplate for your own plugins0.1.0

GHCR

fyos-os/webdesktop shell (nginx)v0.1.3
fyos-os/hubhub as an imagev0.1.3
fyos-agent/openclawagent on the OpenClaw imagev0.1.3

GitHub

FastLocalExpertsengine sources, releasev1.1.0
fle-studiochat, model loader, launch profiles, forgev0.3.0
claude-contexiergesession-notebook plugin for Claude Code (MIT)main

Hugging Face

Qwen3.8-Flash-Next-FLE-GGUFmodel files, MTP head, films1.0
Qwen3.8-Flash-Next-FLE-REAP448-GGUFpruned model, 448 experts1.0

Zenodo

FLE Release-Archivpermanently citableDOI

Timeline

  1. First architecture decisions, including “MCP first”.
  2. Decision to build an own inference engine: FLE.
  3. Rebuild from measured operation: whatever goes unused is removed.
  4. FLE v1.0 released: engine, four models, technical report, eight films.
  5. Contexierge, the session-notebook plugin for Claude Code, released.
  6. Archive DOI on Zenodo. FYOS OS, agent, hub, installer and plugin template v0.1.0 released.
  7. FLE Studio released: a local chat and lab for llama.cpp and FLE, v0.2.0.

← FYOSContact →