NVIDIA Nemotron 3.5 Bolt 74B-A9B
A 74B open-weights agent model squeezed into 24GB — 2-bit quantization, built to run locally on RTX hardware.
About
Nemotron 3.5 Bolt 74B-A9B, released October 15, 2026 (announced October 13), is the local-inference sibling of Super VL. It holds 74 billion parameters in total with about 9 billion active per token, compressed to a 24GB footprint through 2-bit quantization with quantization-aware training. Text, images, and video go in; text comes out, with a 256K-token context and a chat-template thinking-mode toggle.
NVIDIA ships it fully open: weights, training data, and recipes under the OpenMDW-1.1 license, in GGUF and BF16 builds on Hugging Face. It runs through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, with tested targets from a GeForce RTX 5090 PC up through DGX Spark and RTX Spark, plus direct wiring into agent frameworks like Hermes and OpenClaw. NVIDIA positions it for always-on local agents: coding help, monitoring, tool calls, and background tasks without a cloud API bill.
The benchmark table on the Hugging Face card (GPQA Diamond 86.43, BrowseComp 66.75, Terminal-Bench 2.1 50.98, SciCode 41.02) is vendor-run against Qwen3.6-35B-A3B and Gemma-4. Independent benchmarks were not yet available at release, so the 2-bit quality and speed claims still need third-party testing.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported (HF model card): GPQA Diamond 86.43, BrowseComp 66.75, Terminal-Bench 2.1 50.98, SciCode 41.02, HLE 25.30, WebArena (vision) 50.86.
Independent: no third-party rerun at release.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Release facts (October 15, 2026; 74B/9B MoE; IQ2 2-bit quantization; 24GB footprint; OpenMDW-1.1) confirmed by NVIDIA's announcement, the Hugging Face model card, and independent press. Docked for: (1) the benchmark table is vendor-run with no independent rerun at release; (2) the 2-bit compression's real quality cost is untested by third parties; (3) actual deployment memory depends on context length and KV cache, which vendor figures understate.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 8: announced Oct 13, released Oct 15, 2026; GGUF + BF16 on HF; broad local-inference tooling (llama.cpp, vLLM, Ollama, LM Studio)
- · Maintenance 9: weights + training data + recipes published; wired into Hermes/OpenClaw agent frameworks
- · Openness 9: fully open release under OpenMDW-1.1
- · Manual 8: 74B/9B MoE, IQ2 2-bit QAT, 256K context verified on HF card; benchmark table vendor-run, no independent rerun