Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

a local agent now has a memory budget

Meta's Muse Glimmer puts the custody question on consumer hardware: a 30B open-weight agentic model, quantized below 20 GB, with local latency as part of the deal.

Published Updated Read 2 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-08-11

the local agent stopped being a hobbyist promise and acquired a fairly awkward physical requirement: a 30-billion-parameter model that Meta says is meant to run on one consumer GPU, with the quantized package under 20 GB. that puts the personal machine back in the argument, where “owned” has to survive memory budgets and latency rather than just good intentions.

1. a local agent now has a memory budget

sources:

what happened: Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weight model under Apache 2.0. the release is aimed at always-on local agent workflows: coding, function calling, multimodal input, long-context memory, and extended tool use. Meta says quantization brings the model to under 20 GB, leaving room for its KV cache, perception encoder, and speculative-decoding drafter inside a 24 GB or 32 GB machine. the company reports decode-speed gains of 3.1× on an RTX 5090, 1.8× on an M5 Max, and 1.5× on an M4 Max with that drafter.

the collision: “personal ai” becomes a hardware question before it becomes a philosophical one. a model that can stay on your machine without a cloud round trip changes the custody story, but it also gives the owner a new maintenance bill: memory, thermals, quantization choices, model updates, and the small domestic server room hidden inside the word local.

question left open: when does owning the model make a person more capable, and when does it just turn them into the unpaid operator of another software stack?

left on the table

  • OpenSandbox was cut despite strong sandbox and credential-boundary material; it is a broader infrastructure object and would have pulled the edition back toward the recent isolation/control run.
  • Life on the Uncanny Precipice was not used because the primary page was rate-limited during verification, and the available evidence did not justify promoting a voice/attention angle.
  • Determinization in Structure Theories was left as a wildcard: interesting for the hallucination-as-unsupported-canonicalization claim, but submitted in April and too abstract for today’s one-source fieldnote.
  • Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
  • Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup
  • Browser agents — where browser automation helps, where it lies, and how to keep it observable
local-models personal-ai agent-ops