Amir Nurani

@amirnurani

4:15am. Meta Muse just checked me into my flight. AI went from answers to actually running the day. Stay curious.

Beat Saber Song Picker

Displays a list of Beat Saber songs in albums that I own, allowing favoriting, random picking, and random ordering. Also has seven preloaded exercise suggestions, one for each day. This is what I use to help me exercise with Beat Saber after complaining to Muse that Beat Saber lacks a Random button.

Newly submitted builds are featured here. Submit yours

Amir Nurani

@amirnurani

4:15am. Meta Muse just checked me into my flight. AI went from answers to actually running the day. Stay curious.

Your product

Sponsored

Put your logo, a line of copy and an image right here, between the builds Muse developers come to read. Same size as a post.

Cloud Codes runs Unsloth's 2-bit dynamic quant of Muse Glimmer 30B in about 14 GB of laptop memory and tests 100+ autonomous tool calls.

Threadwell Slack is an open-source Muse connector that bridges to Slack's official MCP server via a local OAuth bridge, submitted to the Muse Connector Platform with a full submission pack.

A curated, source-linked catalog of 156 real things people have done with the Muse personal agent, grouped into 16 categories with every entry linked to its original X post.

Been tinkering with speculative decoding on Apple Silicon for a while, and this week I got Meta's new Muse Glimmer 30B working in my project mlx-dspark. On my M4 Pro, the 8-bit model goes from 8.2 tok/s to 18-26 tok/s depending on content. Math is the best case at 3.27x, code 2.5x, chat 2.22x. Output is byte-identical to normal decoding since the target verifies every token, so there's no quality tradeoff; it's just faster. Meta's own DFlash numbers on Mac are 1.5x (M4 Max) / 1.8x (M5 Max), but those are on the 4-bit build, so not really apples-to-apples. 4-bit for me is ~1.7x at ~25 tok/s and only needs ~18GB. The 8-bit run peaks around 40GB, so you want a 48GB Mac for it. Basically, you get 8-bit quality at 4-bit speed. Repo: github.com/ARahim3/mlx-dspark I'm happy to hear feedback, and I'm curious about what other M-series chips get.

Alexandr Wang

@alexandr_wang

1/ releasing muse image today — the first image generation model from MSL. it's agentic: pairs with muse spark to reason through your prompt, search the web, and plan before it generates. people get what they meant on the first try. live now in the Meta AI app.

Bob Troia

@BobTroia

I tasked my @Muse agent with searching every state's unclaimed property database for any money owed to me. Not only did it locate and submit the claims, but it also found some money sitting in an old employer's 401(k) from 1999!

Simon Willison

@simonw

I few notes on Meta's new Muse Glimmer 30B - their first Apache 2.0 licensed open weight model (the Llama models had a janky non-OSI license) simonwillison.net/2026/Aug/10/in…

Built a small bridge so Muse can act on my Mac from my phone, and recorded a real session: it looks at the Terminal, reports the epoch, loss and accuracy it sees, then locks the machine when asked. What struck me building it is how much of the work is permissions, not intelligence: per-action consent, small window captures instead of a live feed, rejecting stale observations before any input. Free beta if anyone wants to try it. Developer here, ask away. try wand here today

Lucas Switzer

@lucas_switzer

A cool thing about @Muse is that every Muse runs on its own Linux machine. So you and your Muse can do normal Linux stuff together — like build your own terminal!

Dmitry Korzhov

@korzhov_dm

holy shit, @Muse just caught $1,742 a week leaking out of my ad account I asked one question about my ad spend. It did everything else by itself: > Pulled Meta, my pixel and GA4 and noticed the numbers didn't match > Split it by age, device and creative, found 4 dead ads eating

Meta for Developers

@MetaforDevs

Muse Image is now available on Meta Model API and priced for production volumes at $0.01/image. It’s an agentic image model that reasons before it renders. Each call searches the web to refine the output through iterative passes and evaluates the prompt for precise elements

No place like localhost has Muse Glimmer implement a 7-stage development plan for a classic arcade game clone, the same challenge earlier run against three Qwen models.

Your product

Sponsored

Put your logo, a line of copy and an image right here, between the builds Muse developers come to read. Same size as a post.

How I AI puts Muse through onboarding, calendar management, goal setting, a one-shot family morning newsletter PDF, browser shopping and the animated avatar.

I'm developing a detailed history book, I'm 333 pages deep, I have used my whole weeks limit and am 8 million deep in my free 1 billion tokens.

Ninus

@NinosMaron

It revived my app that I kept ignoring Have an event app for a specific population and got too tired looking for events and manually enter them in my database My muse is doing that now, and handing my social media for me

Independent benchmarks and analysis of Meta's open-weight Muse Glimmer.

Ran the model with quants (Q4) by Unsloth with latest (build from master) llama.cpp server. It takes ~20GB ram running on M5 Pro with 48GB at about 17t/s. Didn't do any reasoning loops/overthinking. Overall, sits below Qwen3.6 27B, wasn't able to get good code (frontend and backend) results. On the positive side, it didn't fail any tool calls. Your opinions/findings? Watch more: https://www.youtube.com/watch?v=_5wKhkUT438

Built a web-design benchmark for local models and ran Muse Glimmer 30B against Qwen 3.6 27B and DeepSeek V4 Flash 0731.

merve

@mervenoyann

this is super easy to run install llama binary: curl -LsSf llama.app/install.sh | sh run: llama serve -hf meta-models/muse-glimmer-30b --spec-type draft-dflash -fa on --jinja please spread the word

Divine 〽️achine

@Divine_machine

AI assistants are insane time savers, this segment will be massive. @Muse is extremly good for personal things, just today its booked a flight for me, handled emails, got me dinner reservations, and bought tickets to a game, and alerted me for things needing attention. The

Mark Zuckerberg

@finkd

Opening access for developers to build Muse connectors. You bring the API -- Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there. New connectors are live

Matt Johnston's live gauntlet puts Muse Spark 1.2 at 95 and #5 on his board, at $1.25/$4.25 per M tokens and 171 tok/s on OpenRouter; the full bench ran in 17 minutes.

Meta Developers session on what Muse Spark's act-on-perception multimodality unlocks across code, physical action and video workflows, plus Muse Voice Transcribe.

A read-only Muse skill for Partiful: invitations and RSVP statuses, event details, guest lists and mutuals, via community reverse-engineering of Partiful's backend.

Dilmer

@Dilmerv

Hello everyone! 👋🕶️ I put Meta Muse Code + the Unity CLI to the test with a real Unity project to see how far I could push this agentic workflow. Instead of just generating code, I wanted to see how it could actually interact with Unity, run tests, validate changes, port the h

Vals AI

@ValsAI

Meta just released Muse Spark 1.1 and is the new SOTA on MedScribe and TaxEval, taking the top spot from Fable 5 while being 10x cheaper and twice as fast. Meta currently holds the top 2 spots on TaxEval It is also the new #1 on Harvey's Legal Agent Bench, dethroning Grok 4.5

Helicon is an open-source desktop and web app that wraps Meta's Muse Code CLI in a Codex/Claude-style interface, grouping sessions by project with history, resume and diffs on Windows, macOS and WSL2.

Your product

Sponsored

Put your logo, a line of copy and an image right here, between the builds Muse developers come to read. Same size as a post.

Sree

@srikanthvaluri

A realistic 3D model of a BMW S1000RR superbike in Three.js Fable 5.1 VS Muse Spark 1.3

Arena.ai

@arena

Muse Spark 1.1 has entered the Code Arena: Frontend at #9! Muse Spark 1.1 reshapes the cost-performance Pareto Frontier by scoring 1541 at a blended $3.5M ($1.25 per input MToken, $4.25 per output MToken). This is frontier performance at a fraction of the price. Congrats to

Displays a list of Beat Saber songs in albums that I own, allowing favoriting, random picking, and random ordering. Also has seven preloaded exercise suggestions, one for each day. This is what I use to help me exercise with Beat Saber after complaining to Muse that Beat Saber lacks a Random button.

Claire Zhou

@claireszhou

👀 ICYMI: @finkd opened his keynote for Connect this afternoon with a live demo of @Muse Code for 3D asset generation! #MetaConnect2026

Gordon Litefoot

@MrLitefoot

I asked my @Muse for a beat maker this morning. It built this — 16-step sequencer, runs entirely in the browser, free. No code written by me. 🤯 Try it: muse.ai/s/music-librar…

Mr AI Dan

@TheMrAIDan

Two kids. Multiple schools. Swim teams. Emails. Schedule changes. Calendar chaos. So I built EVA, our AI family assistant using Meta Muse. She finds what matters, tracks deadlines, updates our calendar, and answers questions about what’s coming up.

kwindla

@kwindla

Weekend project: play chess with Muse Spark running on a DGX Spark with code Muse Spark wrote. Setup: - projector - HDMI, camera, mic, and display server (Raspberry Pi) - pipecat bot (DGX Spark) - vision pipeline (RTX 5090) - coding agents running on the Pi, the DGX

Arena.ai

@arena

Meta Muse Video just entered the Video Arena at #3. @AIatMeta’s new video model scored 1459 in the Text-to-Video Arena. It outperforms Alibaba’s HappyHorse 1.0 by +30pts and ranks ahead of Grok Imagine, Sora 2 Pro and Google Veo-3.1 models. Meta has now reached the video AI

Meta's official Muse-Glimmer-30B repo: ~29.6B dense model with a 1.8B vision encoder, 131K context, Apache 2.0, with vLLM and SGLang serve commands.

Meta's announcement of Muse Spark 1.3 for Muse Code and the Meta Model API, claiming ~20% fewer tool calls and ~25% fewer tokens than 1.2, with a max reasoning mode.

Bryce Del Rio

@BryceDelRio

Biggest @Muse agi moment. Wife wanted chocolate cake. I wanted to watch football, vs scroll doordash for 20 min. She didnt trust muse. I said lets test it pick out what you want. Sent Quantavius out on my end to find best chocolate cake. They chose the same thing, muse

Been building this for a few months, mostly for myself, and it just got a proper release so figured I'd post it. It's a native GGUF inference runtime with OpenAI/Anthropic-compatible APIs and a chat UI. The whole point is one consumer NVIDIA card + lots of RAM: MoE models that don't fit in VRAM run their experts on the CPU, or split with a hot set cached on the card. It figures out what fits at startup instead of you guessing offload layer counts. Runs Qwen 3.x dense and MoE (incl. Qwen3.8-Flash-Next), DeepSeek-V4-Flash, Ling 3.0, K2-Horizon, Gemma 4, Laguna, Muse Glimmer. Image input via mmproj on the Qwen models. Also does Z-Image-Turbo image gen next to a chat model on the same card. Numbers from my laptop (5070 Ti 12 GB, 60 GB RAM): - Qwen3.8-Flash-Next IQ1_S: ~35 tok/s decode, ~475 tok/s prefill - Qwen3.8-27B IQ2_XXS: ~40 tok/s - DeepSeek-V4-Flash: 6-7 tok/s (that's basically the DRAM bandwidth limit) - Z-Image 1024x1024 in ~15 s with a 35B loaded beside it Stuff I think is neat: - Kernels are compiled at runtime by NVRTC, so no CUDA toolkit in the wheel and no nvcc. Same kernel source compiles as plain C++ for a CPU-only backend. - KV cache in f16 / q8_0 / TurboQu

Your product

Sponsored

Put your logo, a line of copy and an image right here, between the builds Muse developers come to read. Same size as a post.

1079 builds · page 1 of 27