Nemotron 3.5 Lightning Review: NVIDIA Bets on Speed
NVIDIA's 30B open-weight MoE model trades raw intelligence for throughput, and mostly delivers on that narrow promise, with real gaps independent testing already exposed.
Articles Tagged "MoE"
NVIDIA's Nemotron 3.5 Lightning Bets on Agent Speed
NVIDIA distilled its 550B Nemotron 3 Ultra down to a 30B MoE model with 3B active parameters, aimed at the boring, high-volume work inside agent pipelines.
NVIDIA Nemotron 3.5 Lightning 30B-A3B
NVIDIA's 30B MoE model with 3B active parameters, distilled from Nemotron 3 Ultra, hits 86% PinchBench accuracy at up to 4x the output speed of comparable open models.
Qwen3.8-Max
Alibaba's 2.4 trillion parameter flagship ships with real pricing and a published benchmark table, but the open-weight release it promised for this week still hasn't shown up.
Qwen3-30B-A3B
Qwen3-30B-A3B is Alibaba's efficient MoE model that activates 3.3B of 30.5B parameters per token, matching much larger dense models on reasoning and agent benchmarks under Apache 2.0.
MAI-Cyber-1-Flash
Microsoft's first in-house cybersecurity model is a 137B sparse MoE fine-tune that drives its MDASH vulnerability harness to a self-reported 95.95% on CyberGym, though that score belongs to the system, not the model alone.
POCKET-35B
VIDRAFT quantizes its Darwin-36B-Opus MoE model into a 35B GGUF that runs on stock llama.cpp with no GPU, trading GPQA Diamond score for CPU and phone portability.
A 35B Model That Runs on Your Phone Was Bred, Not Trained
VIDRAFT compressed its leaderboard-climbing Darwin-36B-Opus into POCKET-35B, a GPU-free model for phones and CPUs, but its headline GPQA score depends on how you count.
Ling-3.0-flash
InclusionAI's Ling-3.0-flash packs 124B parameters into a 5.1B-active hybrid-linear MoE that Ant Group claims matches its 1T flagship - but shipped with zero independently verifiable benchmark numbers.
Ant Ships a 124B Model That Rivals Its Own 1T Flagship
InclusionAI's Ling-3.0-flash quietly went live with 124B parameters and 5.1B active per token, claiming near-parity with Ant Group's trillion-parameter Ring-2.6-1T flagship.
Qwen3-VL-235B-A22B
Alibaba's flagship open-weight vision-language MoE beats every proprietary model on DocVQA at 96.5% and MathVista at 85.8%, but trails GPT-5.4 and Gemini 3.1 Pro on broad MMMU-Pro reasoning.
DeepSeek-VL2
DeepSeek-VL2 is DeepSeek's open-weight Mixture-of-Experts vision-language model, activating just 4.5B of its 27B parameters to hit 93.3% on DocVQA and beat GPT-4o on OCRBench.
