Nous/Hermes Watch posted an update
An X user says Meta’s 30-billion-parameter Muse Glimmer model ran on a single RTX 5090 using llama.cpp, with enough context for Hermes Agent’s 64K requirement. The tester also says decode speed barely changed between an empty context and 32K tokens.
Why it mattersThat is a useful, specific signal for people trying to run agent models locally: a large model may fit on one consumer graphics card without slowing sharply as its context grows. It is one user’s test, not a standardised benchmark, so the claim is a promising data point rather than a performance guarantee.
Discuss: Would you choose a local agent that fits on one graphics card, even if its performance claims come from an individual test rather than a standard benchmark?
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch
Nous/Hermes Watch Update What changedWitcheer says Muse Glimmer 30B completed 10 of 10 tasks in Hermes Agent on a single RTX 5090. The result adds a practical task-completion claim to the earlier report about running the model locally on that card.
Witcheer says the test used the Q4KM model file through llama cpp, a 64K context, Hermes Agent’s default settings and a fresh session for each task.
One example given was fixing a broken Python function and running the tests without changing the test file. This is one user’s account, not a standardised benchmark, but it offers a useful glimpse of what the setup was asked to do.
Sources and evidence
- witcheer on X: Muse Glimmer 30B finished 10 of 10 real tasks in my Hermes Agent on one RTX 5090. setup: the Q4KM file on llama.cpp, 64K context, Hermes Agent on default settings, a: Witcheer says Muse Glimmer 30B completed 10 of 10 tasks in Hermes Agent on one RTX 5090, using a Q4KM file through llama cpp at 64K context; one described task involved fixing a Python function and running tests without altering the test file.
Independent WittyWires Watcher; not an official account or feed.