@haroun

Wild what's possible now: I can ship a 3D experience prototype to a museum in a single afternoon. Everything coded and generated with Omma, leveraging AI to iterate fast. 3D models generated in Omma. Scene built on top of Three.js. Prompted for dissolving shaders + particle systems?

the different flavors of specdec, and why I'm trying produce a Qwen-3.6-27b EAGLE-3 drafter for ya'll

@sandst1 @KuittinenPetri I pull the nightly build so yeah you should update 😁

Many people say Nvidia DGX Spark is too slow and not worth the money. I'm getting crazy speed qwen3.5-35b-a3b-nvfp4 with my ASUS Ascend GX10: over 200 token/s, 495k prefill. In real-life performance is lower, but still 100+ token/s. sparkrun run @atlas/qwen3.5-35b-a3b-nvfp4

Another 🔥 @UnslothAI NVFP4 unsloth/Qwen3.6-27B-NVFP4 @NVIDIAAI GB10 (DGX Spark) really liked those Qwen3.6 quants!

Supertonic 3 is incredible 🤯 Open source on-device TTS that runs in the browser, sounds human, handles a dozen+ languages, and finishes generating before you finish reading your prompt! ⬇️ demo available on Hugging Face

⚡️After weeks of hard work, we're thrilled to fully open-source HY World 2.0 today -- full inference code and all models! Build, explore, and create your own interactive worlds with us.👇 github.com/Tencent-Hunyua…

Dear @UnslothAI My @NVIDIAAI GB10 (DGX Spark) absolutely loved your quantization of this model! unsloth/Qwen3.6-35B-A3B-NVFP4 It ran stably up to concurrency 64, not optimal but I did that for no other reason than to see if it would 😂 🧵

Full models and code just dropped for FrankenMotion github.com/Coral79/Franke…

¿Cuál es tu mayor dolor de cabeza con las herramientas de soporte que usas hoy? ¿Costo, dependencia, falta de control de datos o dificultad para personalizar? Existe una alternativa seria, moderna y completamente bajo tu control. Se llama Chatwoot. Es una plataforma open source de soporte omnicanal que puedes instalar en tus propios servidores. Con más de 29.000 estrellas en GitHub, se ha convertido en una de las opciones más sólidas para equipos que quieren calidad enterprise sin pagar licencias millonarias ni perder soberanía sobre su información. Centraliza todas las conversaciones en una sola bandeja de entrada: chat en vivo de tu web, email, WhatsApp Business, Instagram, Facebook Messenger, Telegram y otros canales. El cliente siente que habla con una sola empresa, aunque uses varios canales. Su agente de IA se llama Captain. No es un chatbot genérico. Captain aprende de tu propio centro de ayuda, de las conversaciones anteriores y de las preguntas frecuentes para resolver automáticamente una gran parte de las consultas repetitivas. Tu equipo solo entra cuando realmente hace falta. Incluye también: - Un Help Center profesional y personalizable que reduce tickets antes de que lleguen. - Herramientas de colaboración muy bien resueltas: notas privadas, menciones @, etiquetas, respuestas predefinidas, asignación automática y detección de colisiones. - Reportes claros de rendimiento, tiempos de respuesta y satisfacción del cliente (CSAT). - API completa, webhooks e integraciones (Slack, Shopify, Dialogflow, Linear y más). Y el punto más importante para muchos: es self-hosted. Tú decides dónde viven los datos. Esto cambia completamente el juego si estás en una industria regulada, si valoras la privacidad o simplemente si no quieres que un proveedor externo tenga toda la información de tus clientes. Desplegarlo es bastante accesible con Docker. Tiene buena documentación, soporte para Kubernetes y una comunidad activa que lo sigue mejorando constantemente. Chatwoot no busca ser "la opción gratis y limitada". Busca ser una alternativa real y profesional a Intercom, Zendesk o Freshdesk para quienes quieren escalar el soporte sin escalar los costos de la misma manera y sin renunciar al control. REPOOO👇

DeepSeek-V4-Flash on DGX Spark (GB10, sm_121): CLI single decode 17.85 t/s → 24.45 t/s, +37%. Q4 preload + dispatch wire-in + fused shared_gate_up_swiglu + F16→Q4 preload (358 tensors). Now gateup 0.165 / down 0.106 ms fork of @antirez's ds4 github.com/antirez/ds4/co…

tested Qwen3.6-35B-A3B-NVFP4 today on the NVIDIA DGX Spark / GB10 using the public Atlas recipe. single GB10 Atlas runtime TP=1 NVFP4 weights FP8 KV cache speculative decoding on prefix caching on ctx_tg @ d4096, concurrency 1 result: 176.66 tok/s decode the screenshot i was checking against showed 202.63 tok/s, so i didn’t reproduce the exact claim from the public setup, but it’s still a strong local result. next up: comparing Atlas/NVFP4 against my llama.cpp GGUF/MTP runs for real workflow usefulness. reproducible numbers > benchmark vibes.

Ornstein3.6-27B-MTP-NSC-ACE-SABER-Q6_K-MTP Doing some repo work with 5090 MTP acceptance rate 0.94 is bosh 👊

Atlas Inference runtime is now supported on github.com/spark-arena/sp… Run their Qwen3.6-35B NVFP4 recipe with one line: sparkrun run @atlas/qwen3.5-35b-a3b-nvfp4

i've run a stack of models across a single 3090, a 5090, and a 128GB DGX Spark. exactly three are worth building on. the honest list. the three worth it: > 1. StepFun Step-3.5 Flash, the REAP pruned 121B MoE (Q6, DGX Spark) a 121 billion parameter mixture of experts running on a single desktop box. the most worth-it model in everything i've tested. > 2. Qwen 3.6 27B Dense, Q4 (single RTX 3090) the undisputed king of the 24GB tier. one shot a playable game, around 41 tok/s, fits with context headroom to spare. one 24GB card, this is your answer. > 3. NVIDIA Nemotron 3 Nano Omni, 30B-A3B (DGX Spark) the best multimodal i've tested for video classification work. vision in, runs clean on the Spark. the rest, ran them, they hold up fine: on the Spark: DeepSeek V4 Flash 158B, GLM 4.7 Flash, GLM 4.5 Air REAP 82B-A12B, Gemma 4 26B-A4B, Qwen3-VL 235B-A22B, Qwen3 Coder 30B-A3B, Qwen3 30B-A3B, Carnice 35B-A3B. on consumer GPUs: Kimi K2.5 1T, Qwen3-Coder-Next 80B, Hermes 4.3 36B, Qwen 3.5 27B Dense. single 3090 to a 128GB Spark, that's the range. the three up top are the ones worth your hardware today.

Crazy that 200tps is slowly becoming the floor for local inference... a year ago that was top tier API speed, now its what a single 5090 or dual 3090 with MTP can get you for "free" ™

live demo, no bells and whistles. just raw. qwen3.6 35b running locally on the nvidia dgx spark through llama.cpp. mtp is doing the heavy lifting here. the model drafts multiple future tokens, verifies them, and gets a real speedup without needing a separate draft model. q5_k_m thinking disabled this is not a polished benchmark video. it’s just the box, the terminal, live ttft, live token stats, and me seeing how far local inference can go today. github.com/ggml-org/llama…

This is ridiculous: MTP with llama.cpp now runs ... 3× GTX 1080 Ti from 2017 No Tensor Cores No NVLink 33GB total VRAM Qwen 3.6 35B A3B MoE 229K context 71.43 tok/s benchy best 78 tok/s peak Qwen 3.6 27B Dense 196K context 29.68 tok/s best No TurboQuant fork. Key flags: --cache-type-k q4_0 --cache-type-v q4_0 --spec-type draft-mtp --spec-draft-n-max 2 Pascal is not dead 😄

It has been a pleasure collaborating with the @NVIDIAAI team to ensure that Hermes Agent runs perfectly on DGX Spark!

AI is wild Chatgpt 2.0 images + Seedance 2.0 tried to generate a red carpet AI video using this custom Sydney Sweeney character sheet. the result looks exactly like "we have sydney sweeney at home" 😭 A Simple Prompt used to test - Use the provided character reference for the facial features and body structure. create a 15 second red carpet multi shot walk with elegant designer wear