even 30B model is too large to large on local device (low end). meta should provide free hosted model api to use it.
Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
531–540 of 682 posts
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#532Having played around with this model a bit, I am fairly confident that it is not competing in the coding space.
It can do that, but its actual selling point appears to be a different take on guardrails and safety alignment.
Either that or the only new training data left was industrial quantities of dark romance literature and Wattpad.
Clever business move. 131k context is more than enough for that use case, and due to that small K/V footprint, you can probably have a bunch of characters on the same GPU.
Or it's just a happy little accident. We will never know.
___
I was informed that normal people use LLMs for mundane tasks like asking for a pancake recipie.
That it apparently can also do decently.
Unfortunately, it is also very confident, regardless of whether it is actually correct.
So maybe it should actually stay the smut engine and nothing else.
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#533Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#534Pelican, rendered by Muse Glimmer on my Mac running LM Studio (with this model release: https://lmstudio.ai/models/muse-glimmer ): https://tools.simonwillison.net/markdown-svg-renderer#url=ht... It has all of the components of a pelican riding a bicycle, though not exactly arranged in the right order! (For comparison, here are the pelicans I got from Muse Spark 1, 1.1, and 1.2: https://bsky.app/profile/simonwillison.…
> It has all of the components of a pelican riding a bicycle, though not exactly arranged in the right order! Maybe a sign that they didn't have SVG pelicans in the dataset
Picasso's Pelican
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#535The more open weight models get released the greater the market for personal and small business oriented hardware to run these models. This will drive lower cost hardware, which has stagnated in recent years due to most software not needing the performance and capacity.
The opposite happening because foundries are full to capacity making higher margin stuff.
You have to look a little past the current hysteria.
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#536Earlier quoted context omitted.
Based on the benchmarks, it seems that Muse Glimmer barely edges out against Qwen3.6 27B, except for tool-calling skills (MCP, etc.). I wouldn't be surprised if they released it now because they are afraid they wouldn't beat Qwen3.8 27B.
I would hope that Qwen 3.8 is better. It's been 4 months, and we've seen almost no progress in this space. As people have called out, Glimmer appears to be a trade-off rather than a clear winner. And from what I've been reading, no one is expecting Qwen 3.8's model in this space to be a clear winner, but just slightly and marginally better. That's a little concerning as DeepSeek v4 Flash proved at it larger sizes the…
I honestly expect that major advances in the open 30B dense space will take about a year, but expect incremental advances every couple of months from different developers in the meantime.
Qwen 3.6 27B was already a massive gift to smaller homelabs around the world; anything more is just a delightful surprise.
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#537The post suggests that you need an rtx 5090 use it, which is currently selling for around $5,000 USD. I wouldn't exactly call that "my device", since my device costs about 25% of that for the entire computer. For the same cost, you could run on a frontier model on a pro plan for two years. The economics dont make a lot of sense for this to me, so I would love some input on why people want to do this instead (privacy,…
The performance will be so-so, but you can buy an Intel Arc B70 for $1000. There are definitely ways to get going for less.
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#538> Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows. It’s small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation.
The next iteration in LLM products is a 24/7 thinking loop where the claude-code like thing gets input continuously from your wearable, notifications, and newsfeeds and is constantly preparing things for you.
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#539Earlier quoted context omitted.
> On the other hand, one should not discount the value of HN as tastemaker and trendsetter. I would encourage dedicated readers here to aggressively and persistently discount the value of HN as a tastemaker and trendsetter. HN is actually a trailing indicator on tastes and trends, essentially by design. Things only make it to the front page if they get submitted and voted upward by a large number of people. That mean…
So - presumably you have another site in mind that is better. I'd be intrigued to know which one you would recommend.
I don’t. It happens in-person or in private group chats these days.
Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
#540Meta's clearly changing strategies back towards their original "frontier open source", but this time around they have a lot more competition from leading Chinese labs.
I'm all for it though, and I think Glimmer is a fantastic bet on locally-hostable models. I for one would love to self-host as much as I can.