We're having DeepSeek moments every couple of weeks.
Qwen 3.6 hit hard in the self-hosting space. It's incredibly capable for its size, really shaking up what's possible in 64GB or even 32GB of VRAM.
The Prism Bonsai ternary model crams a tremendous amount of capability into 1.75GB.
And, DeepSeek V4 is crazy good for the price. They're charging flash model prices for their top-tier Pro model, which is competitive with the frontier of a few months ago.
The winners in the AI war will be the companies that figure out how to run them efficiently, not the ones that eke out a couple percent better performance on a benchmark while spending ten times as much on inference (though the capability has to be there, I think we're seeing that capability alone isn't a strong moat...there's enough competent competition to insure there's always at least a few options even at the very frontier of capability).
> It's incredibly capable for its size, really shaking up what's possible in 64GB or even 32GB of VRAM.
You can lower that to at least 24GB. I've been running Qwen 3.5 and 3.6 with codex on a 7900 XTX and the long horizon tasks it can handle successfully has been blowing my mind. I would seriously choose running my current local setup over (the SOTA models + ecosystem) of a year ago just based on how productive I can be.
Gonna try it.
We have Qwen 3.6-35b (6) on a 5090 (32GB) and it's blowing me away. Works fine for most (not all) code generation tasks. One developer here has been extremely stubborn about adopting AI; he's finally adopted it, albeit only when it's coming from a local model like this.
DeepSeek V4 Pro likewise is insanely good for the price. I simply point it at large codebases, go get a cup of coffee or browse Hacker News, and then it's done useful work. This was simply not possible with other models without hitting budget problems.
Any chance you'd be willing to talk further about your setup? I have 2 x 3090s in a local machine, and I'm still left with questions about how best to use stuff locally.
You can only run heavily quantized models on all 3/4/5 rtx gpus (with 32gb or less vram) - and you probably want moe versions like Qwen 35b for this to run at speed somewhat comparable to Claude. It’s still not there to be honest but getting there. Personally I mess around with llama.cpp on m5 max with 128gb - it’s a decent setup to try various medium sized things, and runs llms surprisingly well without quantization, at least the moe models.
How is that machine for local inference? It's a serious consideration for me, but getting to hear more from folks that already have it would be helpful.
Two 3090s is 48GB, so it's possible to run the 6-bit quantization comfortably, which is fine. It doesn't start to get notably dumber until lower than that. It won't be as fast as a hosted model, but dual 3090s will be comfortably fast for interactive use with the MoE version and not terrible to use with the dense model. I run the dense model at 8 bits on my dual Radeon V620 desktop machine, which I think would be slower than two 3090s, or at least not notably faster.
Have you done comparisons with 4 bit and seen a noticeable difference for coding tasks?
No, I've just seen benchmarks showing most models start degrading around 4-5 bits. That's not to say they become useless, just that down to about 6-bits (with careful hybrid quantizations like unsloth where some of the layers aren't quantized or are quantized at higher bit depths) the quality isn't measurably degraded, but below that there are measurable differences in performance.
People report good results from DeepSeek V4 Flash at 2 bits (the DwarfStar 4 folks are doing it, and I've tried it on my Strix Halo, but it's too slow to be usable, so I haven't bothered to figure out if it's actually smart enough to use for anything).
Anyway, it's obvious models have to degrade in terms of knowledge, at any quantization, even though it may not show up clearly on benchmarks until lower. If you halve the size of the data available, it necessarily loses information about the world.
The data I've seen is stuff like the KL Divergence comparisons that Unsloth does which show something but not clearly whether there's an observable or significant difference in task performance.
One of the things I'm wondering about is what I'm missing for $LLM to create files on the local FS like Claude and Codex do. What I see instead is stuff just printing to stdout, rather than files on the filesystem.
What am I missing?
You're missing an agent. The model uses tool calls to interact with the filesystem, commands on the system, optionally search (you need a search MCP server, like Brave or Exa, and API key), etc.
I usually use the Zed Agent built into Zed editor for self-hosted models, but you could use Pi, OpenCode, Hermes, Claude Code, etc. there are many, many, agents.
The model just predicts text, Claude Code etc parse the output and do the actual file creation (or run shell commands that do it). If you have Claude Code installed look in ~/.claude/projects/... and you can see the transcripts of your actual sessions, or install Mini-SWE-Agent and play with that to get a feel for what's going on.