Publication with a quote · Models & research · Agents & automation · Infrastructure & self-hosting
Proud to be part of this great release. More to follow on XPU: Qwen3.8-27B attention merge with 16 subgroups per head, bit-identical. Some tweaks for more tok/s. #LocalAI @intel
For fun, and since some of you do robotics: @liquidai LFM3-VL-2B on a single @intel Arc Pro B70, on a work-in-progress branch I’m preparing for #TensorFold. 55 ms to first token on a fresh image, ~270 tok/s for one stream, ~4,500 tok/s served at 32 concurrent streams. And the
Comme dirait une pub sur T’Fou « fais une photo et habille le dragon à ton image ! » n’importe quelle surface s’applique. C’est fou ! Bien joué ! Les modders vont aimer je pense 😁
Publication with a quote · Models & research · Infrastructure & self-hosting
🚀 Proud of this PR merged! Prompt processing for Gemma 4 E4B is now 2-2.6× faster on Intel Arc Pro B70. Massive win for local AI workloads on Intel GPUs: less waiting, more tokens🔥 #OpenSource #LocalAI #IntelArc #Tensorfold
Publication with a quote · Models & research · Infrastructure & self-hosting
Gemma 4 on Intel Arc is now on this version 1.2.0 🎉🎃 Next version will introduce 2x to 2.8x in prefill speed, more than 15k tok/s on the same GPU! @intel @fragtalife #LocalAI
Gemma 4 E4B (NVIDIA Nemotron 3.5 Bolt E4B) now runs on the Intel Arc Pro B70 in #TensorFold main. 4-bit, on one card: 6,772 tok/s prompt, 135.6 tok/s decode. Served chat: 140 / 253 / 426 tok/s at 1 / 2 / 4 streams. Build from main for now; it isn't in a release yet. More prefill
La famille Gemma 4 E4B sur l’Intel Arc B70 ca donnerait dans les 12000 tok/s en prefill à 8k tokens⚡️ Pour faire avaler des tonnes de documents, ça commence à être intéressant 😁 J’attends encore d’autres pistes de perf. Et toujours 1100 tok/s sur 4 flux de code vérifiés bit à
Hi @NVIDIAAIDev @NVIDIAAI , a license question on nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-DFlash: the model card says OpenMDW-1.1, but the GGUF metadata (https://t.co/WiseJnxcot) says nvidia-software-and-model-evaluation-license. Which one applies? We'd like to publish results
Gemma 4 E-series groundwork landed on #TensorFold main: • #1400 E4B geometry in the family config • #1401 tensor plan, BF16/8-bit overrides • #1402 per-layer inputs, bit-identical to MLX bf16 at real E4B shapes Thanks @ashxhart for the sharp reviews 🙏 https://t.co/V1nWpiNciV
Le canard parle français 🦆 #Microduck de @pollenrobotics, en simulation pour l’instant : une phrase, et Nemotron Bolt E4B avec notre LoRA décide en ~74 ms sur une Intel Arc B70 (#TensorFold XPU). Le Mac pilote tout depuis une petite console SwiftUI. Prochaine étape : le même https://t.co/QPVqlqzeSA
An #NVIDIA model, on an #Intel GPU, running a #TensorFold engine, piloting a @pollenrobotics duck. This is either the future of physical AI or the start of a very nerdy sitcom 🦆 Yes, I'm porting Bolt E4B to the B70 now 🐤 https://t.co/O9oCKj693K Coin ! #Microduck
Just ordered two @pollenrobotics #microduck 🦆🦆 (yes, after Jensen’s shout-out, I’m clearly not the only one 😁) Plan once they land: let a local LLM powered by #TensorFold on an Intel B70 decide what they do next. Waddle, turn, nod, sit. It picks one, really fast. No cloud, no https://t.co/[redacted]
Would /v1/decisions (label probabilities from one prompt pass, no generation) be useful to you in #TensorFold’s native engine? Working on Flash Next CUDA (DGX Spark) first; the hook would be shared for others models. @IntelAI Arc B70 users, interested too? cc @ashxhart #LocalAI