Proud of this private custom commission... First and ONLY high-fidelity NVFP4 of `Step-3.5-Flash-PRISM-PRO-NVFP4` (in vLLM!) 1024-sample 512-multi-pass calibration on HQ PRISM Pro dataset — FP4 weights, FP8 E4M3 scales, all 288 MoE experts per layer. Critical heads at full

It's Give Back Fridays: Just gave 208 of you open access to the latest release, foundational level PRISM-PRO Models on hf.co/Ex0bit. Please support the work, ensure long-term progress of local open & free AI for all: ko-fi.com/ex0bit

🫠

Ouch Cloudflare….No Uber Eats?

The Taalas Inc. HC1 is mind-shattering—if you've ever wondered what chatting with a 17K tokens/s ai model feels like, you can now: chatjimmy.ai Can't wait to get api access and give this bad boy a run.

Confirmed – Kimi-K2.5 w/ 50% REAP prune works perfectly, passes all tests (multimodal, multi-tool agent swarms, zero-shot coding, hard math). Should perform well at 75%+ prune and outperform most current models at fewer experts. Still hope for us. Kimi.ai – Any reason not

And there you have it, folks! We have a working 72% prune of my PRISM REAP checkpoint (Ex0bit/Kimi-K2.5-PRISM-REAP-530B-A32B), further pruned by 0xSero — great things happen when we have full community collaboration and support. With a few more tricks up my sleeve — we can

Kimi-K2.5-PRISM-REAP casually giving me my stock projection analysis for the week ahead. Running on Hot Aisle ROCm VM (TP=2, KVM/QEMU - 2x MI300X, ~300 t/s prefill, ~30 t/s decode) – favorite new provider. Go show this team some much-deserved love. (AMD - get these folks more

Interesting FUD. Anthropic hasn’t dropped a single open-source model — can’t say the same of the competition. And calling structured outputs a ‘sophisticated distillation detection method’ is a stretch at best. Extraordinary claims require extraordinary evidence. Keeping an open

Makes me all sorts of happy. Solid numbers. Curious—is this pass@5 or greedy? Would love to peek at the harness.

To date, SSM-Hybrids are still some of my favorites (see: Nemotron-3-Nano-30B-A3B-PRISM). Curiosity bug is in the air, so exploring dynamic selection feasibility eval on Qwen3.5-397B. Based on results, I may release a PRISM-REAP(ER) quant in dynamic safetensor format.

Deep appreciation for foundries that listen to community input. Thank you Qwen — now let's get cooking 🧑‍🍳

RoboSousChefs is in full force. 👀