Repos

262K context with exact sampled MTP on a single workstation card.

Chinese fine-tuned Laya weights for MLX that score message routing, urgency, priority and retrieval keywords in one forward pass of about 27 ms.

Unmodified Laya weights, real NES emulation via Stable Retro, text observations from RAM, model-driven jumping. 27.5 s of game time at 16.79 ms median inference.

Dockerized FastAPI service bundling all three checkpoints behind a language router, with timing-safe API keys, OpenAPI docs, triage and moderation presets and bulk inference.

SGLang patch and kernels for a single mid-range card.

Day-one single Spark build.

A fork tuned for Volta-generation Tesla V100 cards, giving used data-centre GPUs a path to run the 125B MoE. Author is a hobbyist maker.

A reproducible recipe with attributed oMLX and MLX patches, setup scripts and raw benchmarks for the 6-bit quant on an M3 Ultra. Written by an Austin AI consultant for people running big models on Macs.

A homelab owner's practical write-up and configs.

TP2 with expert parallelism and MTP speculation.

MeldLabs' release that runs Flash-Next on Apple silicon laptops at about 45 tokens per second, with a demo video.

Ampere-compatible reader plus benchmark receipts.

Compresses the embedding table so a single Spark holds the whole thing at long context.

Riley Brown paired a ModRetro Chromatic with a native Mac companion so the handheld's buttons and 160x144 screen become a pocket interface for talking to Muse.

Hybrid MoE serving on a 128GB Spark.

Ports ExLlamaV3 kernels to RDNA 3.5 with grouped MoE, int8 mixers and MTP. The author quantizes frontier models down to fit smaller GPUs.

Draft-head speculation and carefully measured Vulkan settings that lift Flash-Next speed on Strix Halo. Notable for publishing the measurement method as well as the numbers.

Plugin that lets opencode use Muse Spark via Meta device login and the monthly Muse Code plan, with no API key needed.

A step-by-step serving recipe for running the 125B MoE on one Ryzen AI MAX+ 395 box, written by a hobbyist. Useful as a replicable baseline for that hardware.

CUDA-graph build for the workstation Blackwell.

MIT engine where an agent writes a short film as code, renders offline and inspects its own frames; most example shorts were made with Opus 5.5 at medium effort.

A roguelike produced in one pass by Opus 5.5, which delegated asset creation to Codex.

.NET 10 demo of GPT-6.1 Sol in Microsoft Foundry using Microsoft.Extensions.AI and the Responses API, covering PR triage, an approval-gated tool call and an evaluation harness.

One typed question per frame; the other paddle is three lines of arithmetic as a control. Physics in WebAssembly, decisions in native Laya over POST /decide.