Suggest a model...
This is my upgraded rig:
- Ryzen9 5950x with 64gb DDR4
- Dual NVIDIA RTX A4000 (16+16GB VRAM)
to whom read my previous posts, i jumped the gun and upgraded my server, it was worthwhile and somewhat cheap given i already had the two GPUs and the DDR4 RAM.
Anyway, i am currently running Qwen3.6-35B-A3B-UD-Q5_K_XL all in VRAM with 65536 context and pretty happy with speed (80-90t/s) and overall responses (mostly chat).
I would like to experiment with something beefier, with CPU offload, that i can run with my llama.cpp. Of course t/s is not a goal here, but precision and accuracy of responses is.
I tried to find a good model with claude and gemini, but always got short. Once the model suggested fully crashed my server (guess fill up RAM and ended up in a swap loop), more then once i ended up chasing non existent models. Pretty annoying.
Considering i would only use between 32 and 48GB or system RAM, can you suggest (preferably with links to HF) some models?
I like qwen3.6, but open to anything.
10 replies
I'm fond of this model. In "thinking" mode
gemma-4-31B-it-uncensored-heretic-BF16.gguf
Edit:
There's also this beast I've been curious about. https://github.com/JustVugg/colibri using deep seek v4 flash. 280B-13b moe.
+1 for llmfan46's heretic variants. I use one of his uncensored Qwen 3.6 35B-A3B variants as my default model.
rpDungeon's Luchador models (gemma derived) are also quite interesting -- they tend to have better prose quality for creative writing tasks.
I've been curious to try experimenting with using Rudo in particular for making more interesting NPC interactions in a text adventure for a while now, but haven't gotten to it yet.
What are you wanting to do with the model?
Agentic development with that setup, you can easily run a good quant of Qwen3.6-27B at full context. unsloth/Qwen3.6-27B-MTP Q5_K_XL and use a fixed template
Creative writing? Probably need a different model. I’ve heard good things about Gemma 4 31B.
Both of those are dense models so won’t be as fast as Qwen3.6-35B-A3B. Also play around with MTP, quants, KV cache quant, KV caching, etc.
The RAM size will limit you on larger models unless you stream from storage.
More chats i guess, maybe also agents in the future, i am keep to explore that but not yet there.
It's funny; I was just reading about someone who went the other way
https://bitworking.org/news/2026/05/surprising-things-i-learned-putting-together-a-home-brain/
At a certain point, it becomes less about parameters and more about tools supporting those parameters. Something like Pithagoras, Understory, MCP tools etc.
https://github.com/thecodacus/pithagoras/
https://github.com/thecodacus/understory
https://www.youtube.com/watch?v=fpvF4n32lsE
https://www.youtube.com/watch?v=IwN-eK1s8og
This is very interesting, i have saved your comment, it feels too soon for my understanding of it all, but both understory and pithagoras feels less obscure than what would have been, to me, only a few weeks ago.
For sure. Feel free to ask questions, too.
The TL;DR I will leave you with is this; some of what we consider as "smarts" in a LLM has traditionally done by brute force - bigger GPU , more parameters.
The alternative approach is to make the llm do less by itself, but instead, call on other tools. That way, you can squeeze out much more from a smaller llm or weaker hardware, so long as the llm is obedient at tool calling.
Think of it like doing arithmetic in your head vs using a calculator. Both provide the answer, but the latter requires much less brain power.
With your amount of RAM+VRAM, I think a "beefier" model would ideally be around 70B, and preferrably MoE since you're talking about CPU offload, but afaik there hasn't been any releases in that size recently. You could try something like Ling 3.0 flash with higher quantization, but I have no idea if it's actually good for chat since I can't run it myself.
But have you tried running a non-MoE model which fits entirely in your VRAM? I don't know exactly how much dual GPU slows down compared to a single GPU with 32GB, but if it's not too bad then you could for example run Qwen3.6-27B which is a step up from 35B-A3B. Then of course Qwen3.8-27B is supposed to be released in a few days.
I would also recommend Gemma 4 QAT Q4_0 which ought to fit pretty snug in 32GB VRAM.
In the same size range there's also Muse Glimmer 30B. It feels like it has some potential, but it seems like you have to be more verbose when instructing it. Gemma 4 has been much better at understanding what I meant from a more open ended prompt.
Based on synthetic benchmarks, it's often claimed that Q5 and Q6 are extremely close to the original model in quality, and that's probably true under certain circumstances (e.g. Q&A). But I think if you're chatting with the model, the "small" quantization errors can accumulate during longer conversations. IIRC there was a study on larger models recently, which showed that while quantized versions got near identical benchmark scores, they spent on average many more tokens to solve the problem. What I'm getting at is that it could be worth testing a model with fewer parameters at Q8, instead of only looking for more parameters.
someone should release something that fits in 96-128gb. Hy 3, mimo, deepseek flash are all close but 256gb (or at least more than 128gb for q4) for high context is requried. Qwen is already at 3.8, but I don't think have updated their smaller models. I assume you've tried q8 version of 3.6, and prefer the larger context window?
Qwen 3.8 27b is about to come out and you should absolutely try that when Unsloth has a Q6_K version.. It will be the best local coding model at that size.
I am very fond of this specific model for 32 GB of VRAM and it will probably be the best agentic thing you can run including your ram.
https://huggingface.co/SC117/Ornith-1.0-35B-MTP-APEX-GGUF
Go into files and pick balanced if you're doing complex agentic work with nuance or pick quality if you want to fit the model and 262k context in VRAM.
If you really want to mess with cpu offloading, try unsloth/Laguna-S-2.1-GGUF at q4_k_m. Its the perfect shape and size to work on your setup. Look into getting dflash working on it. You should be pushing 40tks by my estimate.
Keep an eye out for Ling-3.0-flash-GGUF to be officially supported. It's similar in shape to Laguna but it's supposed to be much smarter