Nvidia publie Nemotron 3.5 Bolt 74B-A9B, MoE 24 Go pour RTX Spark
TL;DR
- Bolt 74B-A9B is distilled from the 120B Super VL parent via Star Elastic compression then 2-bit IQ2_XXS quantization, outperforming 4-bit models of comparable size trained from scratch.
- Bolt E4B ships Oct 23, one week after the Bolt 74B-A9B and RTX Spark retail debut on Oct 16, establishing a hardware-timed model release cadence.
- Nvidia pre-seeded named enterprise deployments at launch: EY for M&A due diligence, Hebbia for credit analysis, Scale AI for Skydance finance agents, AI21 for cybersecurity.
NVIDIA a publié le 15 octobre Nemotron 3.5 Bolt 74B-A9B, un MoE hybride Mamba-Transformer destiné à exécuter « everyday agentic tasks locally on NVIDIA RTX/DGX Spark » dans une empreinte de 24 Go, avec un objectif d'au moins 80 tokens par seconde en décodage.
Le modèle descend du Super 3.5 VL 120B-A12B via Star Elastic, que le rapport présente comme « a push-button, router-guided architecture search and end-to-end compression framework invented at NVIDIA ». Le pipeline ramène les paramètres actifs de 12 à 9 milliards et le top-K des experts de 22 à 9. La quantification est agressive : les couches FC1 et FC2 « constitute the majority of the model weights are quantized to IQ2_XXS at 2.0625 bits per element », tandis que les couches Mamba passent en NVFP4 et que l'attention, les embeddings, le lm_head et les poids MoE latents restent en Q8_0.
Le pari revendiqué par le laboratoire : « the resulting model is better than similarly sized models that are trained from scratch and quantized to 4 bit ». Côté vitesse, le drafter de speculative decoding LiLiCorr, 1,13 Md de paramètres, affiche des longueurs d'acceptation moyennes de 4,68 en greedy et 4,25 en échantillonnage probabiliste. Nvidia mesure 83,99 TPS sur AIPerf SPEED-Bench-Coding et un temps de rejeu de 600,80 secondes sur AA AgentPerf Local, pour un classement #1 PinchBench, #2 WebArena et #2 BrowseComp parmi les « models with a similar memory footprint ».
C'est la seconde pièce de la série 3.5 en deux jours après Nemotron 3.5 Super VL 120B : le gros frère sert ici de donneur, Bolt vise la station locale. Les checkpoints BF16, GGUF et les variantes LiLiCorr sont publiés sur HuggingFace.
Ce qu'en disent les autres médias
-
NVIDIA Official Blog Lire →
First-party confirmation of Bolt E4B release date (Oct 23), full named enterprise customer list, and open dataset plus NeMo Gym RL environment releases on Hugging Face.
Super VL delivers 5x faster throughput than the next best open agentic model in its class.
-
VentureBeat Lire →
Adds post-training benchmark granularity (9-44% gains across five tasks) and flags that the 625 tok/s throughput claim excludes tool-call overhead, reducing its real-world significance.
Post-training improvements range from 9 to 44 percent across five specialized tasks, with medical claims jumping from 28 to 72 percent accuracy after adaptation.
-
NVIDIA Official Blog Lire →
Covers the full software integration layer: Microsoft Execution Containers for sandboxing, hybrid cloud-local routing built into Bolt, GitHub Copilot and Adobe GPU acceleration ready at launch.
RTX Spark is a new beginning for the PC, combining gaming, creative studio and personal AI supercomputer in one device.
-
Gigazine Lire →
Provides the most granular compression breakdown: Star Elastic distillation from 120B followed by IQ2_XXS 2-bit quantization, with 14 benchmark comparisons against competing models.
By applying structured compression and aggressively quantizing to 2 bit, the result outperforms similarly sized models trained from scratch and quantized to 4 bit.
Article original publié par research.nvidia.com
Lire l'article original →Titre original : Nvidia publie Nemotron 3.5 Bolt 74B-A9B compressé en 24 Go pour accompagner le lancement des RTX Spark