DeepSeek V4.1 Flash 1017 everyone's talking about it for a reason. It does great with 3d built this honeybee on a sunflower like it was nothing. Used Deepseek Harness, 1 turn, looks so good!

Super impressive, bro. I'm loving it as well. Check out what I'm building live right now with it.

Is there any way of running this in 128 gigs or less yet?

No, I wish though

I see this quant, is it worth trying?

In your opinion does the deepseek harness do some things better than other harnesses?

Not necessarily, it's just as easy as the others. I use it headless primarily since it has the smallest and quickest footprint out of others I've tried.

Oh this is soo good, look at the details for a single turn.

OG!

Beautiful!

It is legit indeed

This is amazing work from DeepSeek this certainly looks Opus level at more price friendly levels.

Wowoowwowow!!! 🤯

Can ya pass that prompt, would like to give it a go

Would it work with 2x3090+192gb ram?

No unfortunately... heck I can't even get it to work at a reasonable speed with 8x3090s a pro 5000 (48gb) and 256gb of ram.

Wicked! very nice.

Through API or through your colossus?

API wish I could run Deepseek Flash locally :(

I think I might start outsourcing BuddyWorld work to deepseek and leave the core IP to more proprietary American models. I'm out of compute lol. That or I'm going to get my 2nd 20x Claude. I know you're already on 3 😆 😭

@loktar00 Deepseek is very cheap. I use it to supplement Claude

@loktar00 I'm just concerned about losing IP and data to China. And any uneven work, or secret logs, that might come from the weightings. TLDR - I don't full trust it so not sure where to deploy it

@loktar00 I don’t really trust Anthropic or OpenAI either!

@loktar00 They at least have ZDR policies and by enforceable American laws they must follow then

Looks great!

Thanks! Feels weird taking credit but I guess someone had to prompt it 😂

One-shot is a demo. Precise edits across iterations is the test

That’s impressive!! We did 3D whales

So did I actually but I didn't post it because I just wasn't a fan of how it looked😂very "stiff" like these.

It looks just unreal cool

Wow that was fast!

Looks amazing. How fast how much?

@antirez :) DeepSeek-V4.1-Flash-1017 dropped. config.json is byte-identical to V4.1-Flash, same engram layers and token map, same 189GB SSD envelope. it's weights-only, so it's a new download target, not a code change. worth adding? happy to test if no one else already is/has

do the same on opencode. wondering if harness matters

Crazy🔥

Super good

@shhhrrri

deepseek says they tuned it for their own harness, so one shot inside deepseek harness is the best case. the same prompt through claude code is the run i care about

still API only for the big models. a 4090 can't hold the full Flash but a few of the small MLX quants already fit

Okay you're definitely trolling me. My tests are never this good

looks bad

@ojasx1

Nice. What is this ? Web or Blender ?

Full web

It did a great job of the blurring and camera movement.

This detail is absolutely wonderful.

OMG 🙀

Lol this is horrible:))