Full methodology and caveats: github.com/Visual-Synthesize…

MiniMax-M2.7 NVFP4: 127.7 tok/s C=1, 2800 peak C=128 2x RTX PRO 6000 Blackwell, SGLang b12x 0.8.3, TP=2. Prefill 17,286 tok/s @ 8K → 9,908 @ 128K. No spec decode (no drafter yet — when one ships, low-C numbers should jump). Thanks @lukedobl for the quant + kernels.

**All data is public:** - Results & methodology: [github.com/Visual-Synthesize…](github.com/Visual-Synthesize…) - Raw benchmark JSONs: [github.com/Visual-Synthesize…](github.com/Visual-Synthesize…) - 3-run verification data: [run1](github.com/Visual-Synthesize…), [run2](github.com/Visual-Synthesize…), [run3](github.com/Visual-Synthesize…)

198 tok/s on Qwen3.5-122B with 2 GPUs 148 tok/s on M2.5 (229B) 79 tok/s on Qwen3.5 397B (full GGUF) ~15% faster than a WRX90 Threadripper build on the same stack. Stack: 122B NVFP4 — 198 tok/s (SGLang b12x + NEXTN) 27B FP8 — 170 tok/s (DFlash drafter) 229B NVFP4 — 148 tok/s (vLLM b12x) 397B GGUF — 79 tok/s (llama.cpp) Hardware: ASRock Rack B650D4U (AM5) EPYC 4564P (AM5) PLX switch 2x RTX PRO 6000 3-run verified: 197 / 200 / 198 Reproducible. Link in comments.

Congratulations! I'm loving Hermes Agent.

Local AI in 2026 vs 2020. Same energy. ~198x the intelligence. Cables got worse (waiting on shorter cables). Benchmarks got better. 198 tok/s | Qwen3.5-122B | 2 GPUs | no NVLink 148 tok/s | MiniMax M2.5 229B | vLLM b12x 79 tok/s | Qwen3.5 397B | GGUF llama.cpp Build: $2K AM5 ASRock Rack B650 $3K PCIe switch 2x RTX PRO 6000 Blackwell Result: Beats $33K Threadripper builds by ~18% on the same stack Open-weight + local is not just speed. It is: • privacy • data sovereignty • control Some work should never leave the machine.

wow- it looks rad!

yeah its rough, i decided just to stay on my old gen4 threadripper. it already had 256gb ram. its a tiny bit slower for AI inference on blackwell 6000s- but paying for rd-dimm right now is crazy. thats like the cost of another blackwell 6000. also playing around with AM5 servers soon. only 1 gen5pcie slot, but it might not matter for my workflows.

not sure about that. i learned calculus and linear algebra using Khan to build AI.

i was using MJ before it was even named, helped build datasets they used for aesthetic fine tuning- but I haven't touched it in years. Its not a professional art tool. Switch to open models and black box API's.

We’ll see about the timeline. At a certain point it may not matter. For example, with image generation I can already produce commercial-quality work using NanoBanana Pro or open-source models such as Flux2 or Qwen. Raw image quality is comparable. The main difference is workflow control and platform restrictions (censorship). Open-source models perform far better when a project requires a high degree of control or multiple revision cycles. For simple probabilistic generation with minimal creative direction, black-box systems are sufficient. But for professional work, I need deterministic, iterative control loops.

yeah! So true. im really interested in deterministic art systems with generative assistance. constantly inspired by the creativity of the community.

sadly its a cultural issue and a legal issue. Open source will catch up in a few months though so seedance wont really matter for long

anyone saying its nerfed doesnt understand the tech stack. its filtering done pre or post generation for safety and regulation.

its great until its not. ever actually used it on commercial work? the censorship makes it hard to take black box models seriously.

unless you make memes, even closed source seems like a waste of time.

to be fair ive nuked my own inbox a couple times

🔥

🔥🔥🔥 best discord Ive been in since the early Eleutherai 2020 colab days. I learn so much. thank you

very true. ive always found the obsession with realism boring artistically. i think it gimped the art form. replicating a camera or older art styles has always been low hanging fruit and not artistically interesting. Looking forward to watching your video. been a while since i played with AD