AI Trend Research
EN

← October 15, 2026 briefing · added in the 04:00 KST update

P2 Model Open weights Language modelsEfficiency & infraMultimodalFollow-up04:00 updateNVIDIAOllamaHugging Face

NVIDIA officially releases 'Nemotron 3.5 Bolt' 74B-A9B (24GB) and E4B (4GB), with same-day Ollama support

Announced October 15, 2026

Verified (fact)

The previously announced local model 'Nemotron 3.5 Bolt 74B-A9B 24GB' has been released alongside Super VL, along with a new 'Nemotron 3.5 Bolt E4B 4GB' targeting 4GB of memory. A GGUF version of Bolt 74B-A9B and a variant for LiLiCorr speculative decoding were posted on Hugging Face. The 74B-A9B is a vision-language model that fits into 24GB of memory with default 2-bit quantization. Ollama v0.40.5 supports Nemotron 3.5 Super (a 120B hybrid Mamba-2/Transformer MoE with 12B active, C-RADIO vision encoder, up to 1M tokens, and toggleable Thinking mode). According to Ollama's description, Bolt is distilled from Super and uses a bundled drafter for speculative decoding.

Needs fact-check (claim)

NVIDIA claims Bolt 74B-A9B is "smarter than most popular MoE models of similar memory size and faster than most popular dense models."

Why it matters (interpretation)

An agent-oriented vision-language model that runs on a single 24GB consumer GPU landed in local runtimes like Ollama on release day, widening options for local agent setups. Actual quality at 2-bit quantization, however, needs to be verified firsthand.

Sources

Story thread

  1. 2026-10-14 NVIDIA said to release 'Nemotron Bolt 3.5 74B-A9B' on Oct. 15, a MoE model fit into 24GB via 2-bit quantization (community report)
  2. 2026-10-15 NVIDIA officially releases 'Nemotron 3.5 Bolt' 74B-A9B (24GB) and E4B (4GB), with same-day Ollama support

More about NVIDIA