Voxel demos on X are all paint and no thought.
A model renders a pretty pagoda and we call it smart. So I built a challenge instead.
This space is open for sponsorsThe slot is unsold. Support framewire on the donation page, or get in touch to book it. · FRAMEWIREHOUSELead dispatch
Yume_XSep 287
Qwen 3.8 Flash decisively crushes the M5 Ultra with this new recipe by vr8vr8 .
I think I found the fastest dual DGX Spark recipe for Qwen3.8-Flash-Next. 138 tok/s structured, 133 code, 383 structured aggregate at four streams, and now it talks to Hermes. It feels like a
Latest dispatches
qwen38Open any frame for the full clipThis is outcome from 1 sentence +/- "Create a tavern" prompt.
OMP + Qwen 3.8 Next, Hunyuan 3d, Qwen Image 2.1 All operated by agents.
This is getting too good :D
"Make ma a chair" Qwen 3.8 next -> prompt, plan Hunyuan3d -> Create 3d + texture Qwen Image 2.1 -> Rework the texture to real deal quality Hunyuan3d…
Hey Eric, we're working on a pipeline where GPT-6 Astra teaches a Franka arm a new task from a single…
Hey Eric, we're working on a pipeline where GPT-6 Astra teaches a Franka arm a new task from a single sentence in Isaac Sim, and every skill is…
100% free. 100% local.
The Internet v3. Hermes + ComfyUI built the whole site on my DGX Spark. Qwen 3.8 Flash Next for the build.
Running the MiniMax audio and video large model on a single graphics card reduced the time from 6 minutes to 56 seconds!
💥 NVIDIA Research has just open sourced Sol-H3: on a single DGX Spark, it can fully end-to-end generate 768p (1344×768), 5-second, high-definition…
[Special Feature/Gemini 3.8 TTS] Will the voice created by AI disappear in a year?
We compared the voice AI of four companies using the same script on the AI news radio broadcast every morning.
It's been quiet here for a bit.
Behind the scenes though, we've been busy shipping and building. This week: daily updates, one for each day, catching you up on everything.
Every single rtx 5060 ti 16gb vram owner
Bonsai2 27b dense flies on your gpu with the full 262k context window, the mtp head and vision all loaded at once, and it runs at 67 tok/s.
My new local Qwen 3.8 27B setup running on a new M5 max Mac studio.
Powerful machine, and you can run uncensored models. However, is it worth running local AI models?
Qoder is giving Qwen 3.8 Flash away for free
I asked it to build a live stock market dashboard from scratch. It planned the task, built the interface, opened it in a browser, and verified the…
Fully local Qwen 3.8-27B on a 4090.
Settling is done. I ve started running auto researches every night.
Im actively using Qwen 3.8 27B locally and thanks to the universitys system GLM 5.3 Flash Vision is now included in the harness as well.
Pelican Bench is an example of Qwen 3.8 Flash Next that was run using v100 dedicated quantization and…
Pelican Bench is an example of Qwen 3.8 Flash Next that was run using v100 dedicated quantization and kernel, but it is strong that it can be made…
Qwen-3.8-27b on Mac Studio m4 max, amazing results with thinking on, but off isn't bad eiter
Same prompt. Same seed. Thinking on vs off Top: thinking off: 6,071 tokens, 82 s, 76.7 tok/s.
I built a website on the history of the internet.
100% local. Qwen 3.8 Flash Next did the planning and the build. Flux Dev made the images. Hermes was the harness.
Qwen 3.8 Flash Next is fancy.
Whilst people are fawning over Opus 5.5, the local AI community kept cooking.
Vr8vr8 recipe runs Qwen3.8-Flash-Next nearly twice as fast as the mainstream recipes on the same two Sparks. I ran my full grid on it.