Somewhere in Menlo Park, someone at Meta decided the best way to needle the "pay per token forever" crowd was to hand out a genuinely capable AI agent for free, no account required, no cloud bill attached. Rude, honestly. Refreshingly rude.
A 30-Billion-Parameter Model That Fits on Your Gaming Rig
Meta released Muse Glimmer on August 11 — a 30-billion-parameter, open-weight model under the permissive Apache 2.0 license, meaning commercial use, modification, and redistribution are all fair game. Quantized to 4-bit precision, it shrinks to roughly 20GB, small enough to run entirely on a single consumer GPU with 24 to 32GB of VRAM, or on Apple silicon like the M4 Max and M5 Max. No data center, no subscription, no metered tokens.
It's distilled from Muse Spark 1.2, Meta's closed flagship model released just days earlier, using the bigger model as a teacher. Glimmer reads both text and images, holds a 131,000-token context window, and pairs with a smaller "drafter" model (DFlash) for speculative decoding — a trick that roughly triples generation speed on a high-end card like the RTX 5090.
Built to Do Things, Not Just Say Things
This isn't a chatbot with a bigger vocabulary — it's purpose-built for agentic work: planning multi-step tasks, calling external tools, noticing when a tool call fails, and retrying without a human holding its hand. That's the actual frontier right now, and Meta just made a serious version of it available to anyone with a decent GPU and zero cloud budget.
The bigger signal here is strategic: Meta is marking its return to open-source AI after more than a year on the sidelines, and doing it by undercutting the exact business model — metered cloud inference — that most of the industry is betting the farm on. Whether or not Glimmer tops any leaderboard, "runs locally, costs nothing, does real agent work" is a pitch a lot of developers have been waiting to hear.
Free, local, and agentic — Meta just made "rent an AI agent from the cloud" a much harder sell.
If keeping up with all this makes your head spin, that's what we're here for — get in touch.
Source: gHacks Tech News