Nemotron 3.5 Bolt Packs 74B Into 24GB

// 1h agoMODEL RELEASE

Nemotron 3.5 Bolt Packs 74B Into 24GB

NVIDIA’s new 74B-parameter, 9B-active MoE model fits into a 24GB footprint through 2-bit quantization and quantization-aware training. It targets local agents and launches with support across llama.cpp, vLLM, Ollama, Unsloth, and LM Studio.

// ANALYSIS

Bolt makes large local models substantially more practical for developers with a single high-memory GPU. The tradeoff is aggressive quantization, but open weights, training data, recipes, and broad runtime support make this a compelling local-agent release.

–The 9B-active MoE design should deliver better efficiency than a dense 74B model.
–2-bit quantization enables the headline 24GB footprint, though users should expect quality and speed trade-offs.
–Native support across major local inference stacks lowers setup friction considerably.
–Open training data and recipes give developers room to fine-tune specialized agents.
–Super-VL is the better fit for multimodal workloads; Bolt is positioned as the practical general-purpose model.

// TAGS

nemotron-3.5-bolt-74b-a9b-24gbllmopen-weightsquantizationinferencelocal-firstagent

DISCOVERED

1h ago

2026-10-15

PUBLISHED

2h ago

2026-10-15

RELEVANCE

9/ 10

AUTHOR

0x0SojalSec

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL28m ago

Muse Spark Open Releases 816B Weights

Meta released Muse Spark Open, an open-weight multimodal MoE model with 42B active parameters, 816B total parameters, and a 1M-token context window. It targets coding and agentic workflows, with downloads, hosted APIs, and broad inference-engine support available today.

BENCHMARK54m ago

Gemini 4 Argon Matches Astra on ARC-AGI-3

ARC Prize says Google’s Gemini 4 Argon shows a performance profile on ARC-AGI-3 broadly similar to OpenAI’s GPT-6 Astra, putting Google back among frontier contenders. The interactive benchmark tests exploration, memory, goal discovery, and adaptive planning rather than static question answering.

RESEARCH1h ago

AI-Native Air Interface Exposes Physical Backdoor

A Cryptology ePrint paper shows that poisoning just 0.5% of training data can backdoor AI-native 5G-Advanced functions, with a physical trigger requiring only a 2-bit reconfigurable intelligent surface. The attack achieved near-total control over CSI feedback and beam prediction in experiments.