I just claimed local.ai/pawan. Claim your name. local.ai/access?mode=claim&r…

Big thank you to @bleysg for making this possible! Dsv4 running on a single dgx spark!! My hermes agent thanks you! Here is what he sent as a validation :) • Model: DeepSeek-V4-Flash (0731), 284B MoE / 13B active • Quant: IQ2XXS, 87GB on disk • Host: single DGX Spark, GB10, 128GB unified • Context: 262,144 (installer default is 32,768) • Engine: DwarfStar 4 (antirez) via Entrpi's Spark fork, v0.5.1 Context ceiling test Stepped 69,632 → 131,072 → 262,144 → 1,048,576. All started in ~20s at identical memory, which was suspicious — so verified with a real payload: prompt_tokens: 124,221 elapsed: 129s prefill: ~963 tok/s reply: correct Confirms the ~1,000 tok/s prefill claim independently. Generation speed at 262K (client-side, includes TTFT) 10.6 / 14.2 / 14.1 tok/s Speculative decode — improved at longer context accept rate: 77.1% → 85–89% tok/step: 3.1 → 3.6–3.96 Memory 113–115GB / 121GB used, ~7GB available. Stable across all context sizes. Zero OOM. Settled on 262,144 — 8× the installer default, verified stable, matches @MiaAI_lab independent default.

5/ Quality check — asked it to explain confidential computing: "Confidential computing utilizes hardware-enforced TEEs, such as Intel SGX or AMD SEV, to create isolated execution domains where code and data are encrypted at rest, in use, and in transit." From a 6 GB model. Nailed the threat model too.

4/ Results: • CPU: ~7.5 tok/s • CUDA: ~28 tok/s • 3.7x speedup • Model memory: ~6 GB • Vision, tool calling, and thinking mode all work

3/ The fix: clone PrismML's llama.cpp fork (prism branch), build with cmake targeting sm_121 (Blackwell compute 12.1), CUDA 13.0. cmake .. -DGGML_CUDA=ON -DGGML_CUDA_F16=ON -DCMAKE_CUDA_ARCHITECTURES=121 ~15 min build on the Grace CPU.

2/ PrismML ships pre-built llama-server binaries, but CUDA builds are x64-only. The DGX Spark is ARM64 (Grace + Blackwell). So setup falls back to CPU → 7.5 tok/s. Barely usable.

I was hesitant to share this, but wanted to put it out there in case folks are looking for inspiration. My mother passed away 2 years ago from cancer. I'd been wanting to write a cookbook of her recipes Telugu home cooking, the dishes I grew up with in her memory and also as a fundraiser for cancer research. But between two young kids and work, it never happened.A $20 @OpenAI subscription and @steipete's @openclaw gave me a thinking partner that kept me on track, helped format everything, and reminded me of deadlines. . If you are hesitant to try starting that project you've been putting off... some of these tools really do make the impossible feel possible. Give it a try !

hey @MichaelBarbaro just noticed The Daily episode titles are fully on the em dash train now. Thought we had an understanding 🫡 🤣

I wrote a cookbook for my mother. She passed away from cancer two years ago. This publishes on her anniversary. 29 Telugu home recipes organized by memory, not meal type. The simple food of an Andhra kitchen. PDF is free. Also, on Kindle pawankhandavilli.com/ammas-k… All royalites go to cancer research, or you can support directly in her memory. All proceeds go to @ConquerCancerFd If you have a recipe or memory from your own kitchen, I would love to read them. [email protected]

Big +1 to this ! Please @tim_cook 🙏