LLMs I Use
Oh god oh fuck I like LLMs now
A list of the various LLMs that are useful to me. Ideally changes over time.
Local
I have a consumer GPU capable of running models, so I run a few at home. The two main ones that see use are listed here, even if I may have more/less experimentally.
The specific card is a 7900XTX; AMD’s flagship from the era. 24 gigabytes of VRAM lets me run up to ~35B models at reasonable quantizations and contexts without CPU offload.
I use llama.cpp as my inference engine, compiled from source with a small bevy of flags.
Muse Glimmer
A ~30B parameter model with a really good vision tower, and supposedly-excellent local agentic performance. The full 131kT context window fits onto my GPU just fine, and I get to enjoy it at quite a good generation speed.
In terms of functionality, I’d rate it quite high. It’s very generalist, and has a decent voice. It’s competent at most tasks, especially when aided by web search, grounding, and occasional guidance/advice.
Qwen 3.8
A ~27B parameter model with built-in multi-token prediction. I run it at 131kT context window again, which gets me more than enough for daily use. I don’t actually use the base Qwen 3.8, rather an ablated faster-reasoning fine-tune.
It’s very capable when it comes to code, but it seems generally benchmaxxed for that. I wouldn’t trust it to be a generalist, even though it might be capable of it.
Pay-As-You-Go
For PAYG models, I use both NanoGPT and Opper. The latter is a shared account, the former is mine alone. NanoGPT has access to a wider variety of models, but at times has flakier service than Opper.
MiMo V2.6
I’ve primarily used Flash, though Pro is good too. Has quite a strong recency bias and problems with tool-calling.
A very competent fast model that lets me fuck around a lot. It has an excellent voice and lacks obvious slop/Claudeisms. In a word, refined.
Its downsides are on the inference side of things. With aggressive classifiers/filters and broken tool-calls, it’s likely an inference stack that hasn’t quite caught up to the model it’s trying to serve.
I use Flash for my personal agent Mikiru.
Subscriptions
I only have two subscriptions at the moment; NanoGPT and OpenAI (through my work).
GPT-5.6 Terra
Primarily used for infrastructure management, I find it quite competent in general as well. If prompted well, it lacks the usual sloppy vibes of weaker GPT-5.Xs, but still has a tone to it.
This model runs Bryl, my more public agent. I have effectively unlimited Terra with the subscription I’m provided.
GPT-5.6 Sol
In DMs with Bryl. Really good model, and has a genuinely really nice voice at least with Bryl’s prompting. I can’t speak to its general capabilities, as I avoid using it in the chatgpt UIs/coding tools due to its usage consumption.