Resources

MarkTechPost summarizes Meta's numbers: 75.4 on DeepSWE v1.1 (Opus 5 74.0, GPT-5.6 Sol 72.7), 88.8 on Terminal-Bench 2.1, and 98.1 on MRCR v2 at 512K–1M context.

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.

Meta's launch post for Muse Glimmer, an Apache 2.0 30B model for local agents that fits in ~20GB at 4-bit and runs on M4/M5 Max Macs, RTX 5090s or 24–32GB GPUs.

TryAI has frontier and open models, including Muse Spark, each build a raycaster, a Rubik's cube, a calculator and a fourth app, and compares the results.

Decensored Muse Glimmer 30B made with Heretic v1.4.0, shipped with a reproduce directory.

sequelbox's Tachibana-Agent fine-tune of Muse Glimmer 30B, part of a series also released for Gemma 4 12B and Qwen3.6 27B.

Wavect's guide covers Glimmer's hardware targets (24–32GB) and tool-use scores such as MCP-Atlas 75.5 and SWE-Bench Pro 51.2, and recommends a 20–30 task pilot before production.

Codus reads the three benchmark charts Meta published for Muse Code and notes Claude Opus 5 wins all three, including Meta's own internal eval.

Hugging Face's launch post covers day-0 transformers, llama.cpp and vLLM support, Inference Endpoints, speculative decoding, TRL fine-tuning and agent demos for Muse Glimmer.

Meta's developer post on running Muse Glimmer on a single consumer GPU with vLLM, llama.cpp and ExecuTorch, with quantized builds in 24-32 GB and cookbook recipes.

Experimental OpenVINO INT4 conversion of Muse Glimmer 30B that needs development builds of Optimum Intel and OpenVINO.

Codilar explained that Shopify products became searchable and purchasable inside Muse through Shop Pay at launch with no merchant setup, and what product-data and checkout readiness merchants need for the new channel.

A DPO fine-tune of Muse Glimmer focused on prose quality and humanlike characters, trained on roughly one-third writing and two-thirds roleplay data.

A Jacobian lens (J-lens) for interpretability on Muse Glimmer 30B's 52 layers, reading out which tokens each activation is poised to verbalize.

QAInsights' hands-on first look at Muse Code covering CLI commands, syntax and setup on launch day.

Simon Willison ran an 18.16GB build of Muse Glimmer locally, testing code exploration and image description. He found multi-step reasoning and tool use strong and creative generation mixed.

Red Hat AI's FP4 weight and activation quantized Muse Glimmer 30B.

NVFP4 W4A4 quantized weights of Muse Glimmer 30B.

Meta's own GGUF conversions of Muse Glimmer 30B for llama.cpp.

Meta's launch post for Muse, a personal agent powered by Muse Spark 1.3 that runs in a dedicated Muse Secure VM and acts across email, travel, forms and purchases.

A suite of 27 GGUF quantizations of an abliterated Muse Glimmer 30B produced with Heretic v1.4.0.

DFlash 2 speculative-decoding draft model for Muse Glimmer 30B from z-lab, run inside a speculative decoding server alongside the target model.

A quantization-aware-trained Q4_0 GGUF of Muse Glimmer 30B for llama.cpp. On held-out tokens it measures closer to BF16 than Meta's official Q4_K_M: 0.0213 vs 0.0228 KL and 95.9% vs 95.6% top-token agreement.

Muse Glimmer 30B abliterated with Heretic v1.4.0 using self-organizing maps and magnitude-preserving orthogonal ablation.

DataCamp's Josep Ferrer ran Muse Spark 1.3 on three real coding tasks. Two used 23–32% fewer completion tokens, but a refactor used 70% more, for a net 12% cost increase.

Fireworks made Muse Glimmer 30B available on launch day, pitching high-concurrency, cost-effective serving for always-on agents.

Sediment's fine-tune of Muse Glimmer 30B that answers factual questions when confident and says "I don't know" otherwise, calibrated for the AA-Omniscience setting.

Stark Insider introduced Muse to its own AI agent setup to test whether Muse's one-minute onboarding costs depth compared with OpenClaw.

Wavect's buyer guide to Muse Code pricing, training-data rights, client repositories, agency margins and routing workloads safely.

A 4-byte GGUF header edit that raises Muse Glimmer 30B's 131,072 context limit to 262K. The author shows YaRN scaling only adds prefill cost on this architecture, and verified 262K with 8 slots on one RTX 4090.

Muse Glimmer 30B quantized with Intel AutoRound at a 3.5-bit target and packed with llm-compressor, tested on vLLM.

AWQ INT4 quant of Muse Glimmer 30B calibrated on STEM and agentic data across ten languages, 24.02 GB.

turboderp's self-calibrated EXL3 quants of Muse Glimmer 30B down to 1.75 bits per weight, with calibration and eval traces.

Simon Willison reads the Muse Code launch as more evidence that long-sequence agentic tool calling is now the key model trait.

explainx.ai breaks down Meta's Sep 19 opening of the Muse connector platform: developers bring the API, Muse brings the agent, browser and user context.

Abliterated Muse Glimmer 30B GGUF quant ladder, updated with a 1.63 GB abliterated DFlash drafter and a 1.40 GB multimodal projector.

Rohan Adwankar found Muse runs on Cloud Hypervisor microVMs with a 327 MB Rust harness that boots in ~40 s, an on-VM Postgres with 194 tables and 60+ privsep tool workers.

GeekWire reports Amazon blocked Muse from making purchases as of September 20, saying the agent accessed its store without permission, didn't identify itself, and appeared to store credentials.

PyTorch added end-to-end Muse Glimmer support to ExecuTorch; on an M5 Pro, DFlash speculative decoding lifts image+text decode from 21.6 to 33.0 tok/s, and it powers the Pi coding agent locally.

llama.cpp imatrix quantizations of Muse Glimmer 30B with image support via mmproj and MTP/DFlash notes.