Noteworthy Models 2026 (22B-36B)
Collection
Interesting models in 2026 that can run decently with 24GB of VRAM (or a lot of patience) • 7 items • Updated • 1
EsKa
SerialKicked
AI & ML interests
Working on Lethe AI, a modular, object‑oriented C# library that connects local or remote LLM backends to applications. Feature long term memory, backend agnostic code, web search support, and agentic tasks, all rolled into one neat C# lib.
Recent Activity
updated a collection about 4 hours ago
Noteworthy Models 2026 (22B-36B) updated a collection about 4 hours ago
Noteworthy Models 2026 (22B-36B) updated a collection about 7 hours ago
Untested / Unsorted modelsOrganizations
updated a
collection about 4 hours ago
updated a
collection about 7 hours ago
liked a
model about 7 hours ago
updated a
model 1 day ago
updated a
collection 3 days ago
Le Chonk- Leaky Homunculus by the French Budget-Burning Committee
3
#10 opened 4 days ago
by 
AdrienneNoctis
Usual report I guess :)
❤️ 1
2
#1 opened 4 days ago
by SerialKicked
updated a
collection 4 days ago
liked a
model 4 days ago
Will you create a similar version for Qwen 3.8 Flash Next?
9
#3 opened about 2 months ago
by 
Hector111
Thulsa Doom
🚀😎 4
129
#18 opened 3 months ago
by 
jukofyork
Am I doing something wrong?
43
#1 opened 17 days ago
by 
Atmaniac
2026 Model Comparison: Complex image recognition and painting
❤️🔥 3
12
#4 opened 3 months ago
by SerialKicked
liked a
model 24 days ago
Project Solo
👍 1
32
#1 opened 2 months ago
by 
fuzzy-mittenz
Nice work!
❤️ 3
2
#5 opened 3 months ago
by 
sophosympatheia
reacted to Undi95's post with 🔥 30 days ago
Post
4,665
Yo, I'm back, and I'm currently trying to teach a local LLM to stop waiting for a prompt kek.
I'm building a small proof of concept: can an open-weight model (Qwen3.8-27B, running locally on 2 RTX 5090 GPUs) learn to direct itself, then improve from its own exploration, without a human in the loop and without breaking it for normal use?
No user, no task. The model only gets observations from its environment. Each turn, it writes its own agenda (goal/open questions/next step), then picks an action: search the web, read a page, or take a note.
The environment is the judge, not another LLM. A note is accepted only if it quotes the page it read word for word. Facts are checked by exact match.
Later, code will be checked by actually running tests.
The best episodes become fine-tuning data (LoRA). The helper system prompt is removed at training time, so the behavior has to live in the weights.
Each new model goes through a fixed benchmark gate: math, general knowledge, "does it still answer humans normally?", autonomy, and learned facts on held-out sources. It's kept only if nothing regresses, otherwise it's discarded. Then the loop starts again.
The full pipeline works end to end: collect, train, merge, deploy, benchmark. The baseline is clear. Without any instructions, the base model's real autonomy is zero: it behaves like a chatbot waiting for a question. That's the number this small project is trying to move.
I haven't found a public tool that runs this whole loop (self-directed exploration, verifiable rewards, continual fine-tuning and a regression gate) on home hardware. The goal isn't AGI in a bedroom. It's to show that anyone can try it, measure it honestly, and see where it breaks.
Code and results will be released once the first real iterations are done. At the moment the code is... running, but made with scotch and stick, still only a PoC I want to try.
Did you already tried something like that? What was your result? I'm curious!
I'm building a small proof of concept: can an open-weight model (Qwen3.8-27B, running locally on 2 RTX 5090 GPUs) learn to direct itself, then improve from its own exploration, without a human in the loop and without breaking it for normal use?
No user, no task. The model only gets observations from its environment. Each turn, it writes its own agenda (goal/open questions/next step), then picks an action: search the web, read a page, or take a note.
The environment is the judge, not another LLM. A note is accepted only if it quotes the page it read word for word. Facts are checked by exact match.
Later, code will be checked by actually running tests.
The best episodes become fine-tuning data (LoRA). The helper system prompt is removed at training time, so the behavior has to live in the weights.
Each new model goes through a fixed benchmark gate: math, general knowledge, "does it still answer humans normally?", autonomy, and learned facts on held-out sources. It's kept only if nothing regresses, otherwise it's discarded. Then the loop starts again.
The full pipeline works end to end: collect, train, merge, deploy, benchmark. The baseline is clear. Without any instructions, the base model's real autonomy is zero: it behaves like a chatbot waiting for a question. That's the number this small project is trying to move.
I haven't found a public tool that runs this whole loop (self-directed exploration, verifiable rewards, continual fine-tuning and a regression gate) on home hardware. The goal isn't AGI in a bedroom. It's to show that anyone can try it, measure it honestly, and see where it breaks.
Code and results will be released once the first real iterations are done. At the moment the code is... running, but made with scotch and stick, still only a PoC I want to try.
Did you already tried something like that? What was your result? I'm curious!
- 23 replies
@Undi95 Welcome back! That's the state of HF, nowadays. What's not full-on bots are a bunch of imbeciles who have delegated the little thinking they used to do to machines (and here, the two paired together). At least they are very easy to spot.
A side-effect of making the tech easy to use by those "people", is that it made HF incredibly cancerous. Just block/report/delete and move on, like everyone else does.
liked a
model about 1 month ago
