Mahmoud Gamal El Deen @righteous_ai
Egyptian Muslim engineer. I run big models on single GPUs and post the real numbers. No cloud. No hype. Just local. Hurghada Joined October 2026-
Tweets14
-
Followers3
-
Following65
-
Likes1
@geldeki @coldniko Makes sense, 280W per card is a sweet spot for 3090s. Have you tried the newer Strata engine builds on that setup yet? Curious if your cold prefill moved past 3K tok/s.
@lyesbahi @coldniko Useful to see the per-user number at 2. Single-stream NVFP4/FP4 on my 5090 for comparison: 157.5 tok/s avg decode, but with a 187,860-token prompt (73% of 256K), so that's deep-context decode, not a short prompt. Have you run IQ3_S that deep? x.com/righteous_ai/s…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
@seanhighness Good jump. NVFP4/FP4 data point on a similar box (5090, 128GB DDR5-5200, Core Ultra 7 265K): 6,840 tok/s prefill at 187,860 tokens (7,014 live peak), 161.7 tok/s peak decode. Curious how IQ3_S prefill scales past 90K on yours. x.com/righteous_ai/s…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
@loktar00 Worth it. On one 5090 + 128GB DDR5, current Strata gets Qwen3.8-Flash-Next NVFP4 to 6,840 tok/s prefill and 157.5 tok/s avg decode on a 187,860-token prompt. Reading ~188K tokens takes under 30s. x.com/righteous_ai/s…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
@MythomancerX @robertocarrizo It holds up. Same card, NVFP4 on Strata with 128GB DDR5: 6,840 tok/s prefill on a 187,860-token prompt, then 157.5 tok/s avg decode. 9,074 hot experts in VRAM, the rest in pinned RAM over PCIe Gen5. Monitor screenshots: x.com/i/status/21139…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
@BottleCapAI Nice. Fewer thinking tokens stack with local decode speed: at ~157 tok/s on a single RTX 5090 (NVFP4, Strata), a 38.5% cut is real wall-clock saved on long agent runs. Any plans for an NVFP4/FP4 build, or GGUF only for now? x.com/righteous_ai/s…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
@coldniko @mhelcig Huge. Strata has been great for Qwen3.8-Flash-Next here: one RTX 5090 + 128GB DDR5, NVFP4, 161.7 tok/s peak decode at 187K context, with the cold experts in pinned RAM over PCIe Gen5. Keen to run DeepSeek on the same box and share numbers. x.com/righteous_ai/s…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
@ItsmeAjayKV The 6B-active design is why it runs so well locally: on one RTX 5090 + 128GB DDR5 with Strata, NVFP4/FP4 held 157.5 tok/s avg decode (161.7 peak) on a 187,860 prompt, with 9,074 experts cached in VRAM. Hard to beat that quality/speed trade at this size x.com/righteous_ai/s…
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
I post measured local LLM runs: real throughput, engine head-to-heads, copy-paste recipes. Follow if that's your thing. @Alibaba_Qwen
Full recipe (flags, systemd units, logs): github.com/Righteous-AI/S… What should I run next? Drop a model + context length below 👇
161.7 tok/s decode. 188K-token prompt. One RTX 5090. A frontier MoE (Qwen3.8-Flash-Next, NVFP4) running fully local on Strata by @coldniko. Real numbers, straight off the live monitor 🧵👇
Erman Eroglu @geldeki
514 Followers 2K Following Building software by day. Running local models at night. AI Researcher 🔬 AI/Agent builder 🤖 Indie Hacker 🚀
Andrej Karpathy @karpathy
4.4M Followers 1K Following
Abdellatif @Abdella6if
11K Followers 212 Following Building qaf and agentset. Previously co-founder of tarteel and an engineer at twitter. 🇵🇸
Teknium 🪽 @Teknium
135K Followers 6K Following Cofounder and Lead Engineer - Hermes Agent @NousResearch, prev @StabilityAI Github: https://t.co/LZwHTUFwPq HuggingFace: https://t.co/sN2FFU8PVE
Erman Eroglu @geldeki
514 Followers 2K Following Building software by day. Running local models at night. AI Researcher 🔬 AI/Agent builder 🤖 Indie Hacker 🚀
Stable Diffusion Tuto... @SD_Tutorial
13K Followers 91 Following 👉 Ai models local installation 👉 Comfy Workflows 👉 Tutorials (Image Gen, Video gen) FOLLOW WEBSITE 👇👇
FP4 Brain @thefp4brain
523 Followers 502 Following AI researcher ◆ post-training ◆ bare metal ◆ dev ◆ neovim + tmux ◆ building frameworks that help make sense of the future of software development
Mia @MiaAI_lab
43K Followers 366 Following Building with AI & LLMs // Running frontier open models locally
John Carmack @ID_AA_Carmack
4.2M Followers 297 Following AGI at Keen Technologies, former CTO Oculus VR, Founder Id Software and Armadillo Aerospace
Nous Research @NousResearch
288K Followers 31 Following A bunch of nerds making progress https://t.co/vrD0aDIGDQ
Black Forest Labs @bfl_ai
56K Followers 0 Following Powering visual intelligence everywhere. Creators of FLUX models. join us: https://t.co/XiNaRNYjFH
Mo Elgaraihy @EngMoElgaraihy
47K Followers 368 Following 🇪🇬🇺🇸🇸🇦 Software Engineer | SaaS Builder | AI Content Creator (YouTube) 📩 Biz: [email protected]
Mohamed Moussa @themmmoussa
2K Followers 104 Following CTO at @tarteelAI. ☪️️ 🇨🇦 engineer working at the intersection of "I wish there was" and "that's impossible".
KVCache.AI @KVCache_AI
1K Followers 109 Following Hi, this is https://t.co/EO7MXLjjSU official account. We build systems for efficient LLM serving, including KTransformers, Mooncake and AgentENV.
Unsloth AI @UnslothAI
102K Followers 483 Following Run and train models locally with the Unsloth Desktop app. 🦥 https://t.co/2kXqhhvdCD
Tech2Wild @Tech2Wild
5K Followers 223 Following 🎮 Tech, gaming, AI, and everything in between. 🤖 Building with it, not just talking about it. 🔥 From the mind of @ToNYD2WiLD
Wayne Lowry @wikiwayne
2K Followers 3K Following I test AI so you don't have to: models, agents, coding tools, hardware. Real runs, real costs, honest results. Call me Super Intelligence.
HUMAIN @HUMAIN
51K Followers 8 Following Empowering the future of AI, from leading AI infrastructure, to advanced intelligence. A Global AI Powerhouse.
Dyad ✨ local, open-... @dyad_sh
1K Followers 325 Following Local Bolt/Lovable/v0 alternative ✨ Free download https://t.co/Mg6TSdscwZ 2000+ builders: https://t.co/zi7m6ncdwZ Open-source: https://t.co/tFUt1MIcRY
Will Chen @wwwillchen
73 Followers 61 Following Building @dyad_sh - local, open-source AI app builder. #BuildInPublic
DHH @dhh
940K Followers 492 Following Father of three, Creator of Ruby on Rails + Omarchy, Co-owner & CTO of 37signals, Shopify director, NYT best-selling author, and Le Mans 24h class-winner.
Tom Ballard @tcballard
4K Followers 2K Following Building things I wanted. Open sourcing them in case you wanted them too.
RyanLee @RyanLeeMiniMax
12K Followers 359 Following Head of DevRel @MiniMax_AI. Building @MiniMaxAgent and @Hailuo_AI. Make more model be open
CV.YH @0xCVYH
19K Followers 4K Following Caio Vicentino on YT. @YieldHackerPass - @gotasofficial - https://t.co/XpTkcm6ZQp - https://t.co/rqT09YELjb - https://t.co/jEmSuZfVOw - https://t.co/HcjE7kxFFA
mr_r0b0t @mr_r0b0t
12K Followers 8K Following https://t.co/0hLf76cK36 https://t.co/hJPF2vCxXx https://t.co/6IbAvxL4xQ observer of a human on a rock somewhere in space
TechMD @TechMDAI
3K Followers 4K Following 🔬 MD fueled by a deep passion for Medicine & Tech. 🌐 Exploring the frontiers of VR/AR/MR & BCI. 🤖 AI enthusiast.
Marc Seal | Aevonix R... @Kurcide
6K Followers 1K Following Founder, Aevonix | 36x DGX Spark Cluster | Building local SI agents with A.E.V.A. & Protagine | Sharing experiments, code & lessons from my Sparks & 6000 Pros
iTarek @iTarek
22K Followers 938 Following MIMV LLC. CEO, Founder of https://t.co/OF2b1WkaP6. - مدير شركة أم.أي.أم.في مؤسس آي-فون إسلام, مطور ومدير مشاريع وقدمت العديد من التطبيقات الشهيرة | @iphoneislam
Tianyi Cui @tianyi
55K Followers 486 Following Work on @DeepSeekHarness and other projects at @deepseek_ai Product Division. DM is open. Opinions are my own.
الموسوعة ال... @QuranPediaApp
66K Followers 0 Following فوائد من التفسير وعلوم القرآن، وخدمات تعين على القراءة والفهم والحفظ والبحث، وجديد الموسوعة وتحديثاتها. https://t.co/5VRkFOR2xM
Ash Hart @ashxhart
9K Followers 495 Following Doing my bit to bring local AI to everyone | building https://t.co/H2dBxCSaUV | https://t.co/E891qTiY1L | https://t.co/EHDAxlBLUd | Thoughts are my own.
netrunner @plotarmordev
8K Followers 773 Following Cofounder & engineer. Running local AI, building things with it, and sharing what I learn.
shaun @Biggest
4K Followers 1K Following open weight is the only way Enthusiastic about AI! Whee! leaky faucet message me! prev @MetaAI, @AMD, @ServiceNow
Blackfrost_SI @Blackfrost_AI
8K Followers 719 Following Leader in Super Intelligence, S.I. startup specializing in cybersecurity and custom security models. ❄PRO open source
nico @nicochristie
33K Followers 6K Following exploring @offpolicyinc. prev creator of @tryshortcutai
Need More VRAM @needmorevram
811 Followers 67 Following AI Engineer • running models locally until something catches fire llama.cpp • vLLM • GGUF • homelab
Miguel Ángel Durán @midudev
344K Followers 373 Following 🧠 Enseño Programación e Inteligencia Artificial 👨💻 Software Engineer y Speaker ⭐ +18 años de experiencia · +3M comunidad
You might like
