On 2026-10-15 NVIDIA released Nemotron 3.5 Super VL, an open multimodal agent model with 120B total / 12B active parameters: a Mamba-2–Transformer hybrid MoE with a C-RADIO vision encoder, text/image/document/video input, up to 1M context, and built-in MTP plus LiLiCorr speculative decoding. Weights, datasets and recipes ship under OpenMDW-1.1 on Hugging Face, OpenRouter, ModelScope and NIM. Alongside it, NVIDIA published Bolt 74B-A9B 24GB, a natively 2-bit (QAT) local agent model for RTX/DGX Spark; the 4GB Bolt E4B follows on Oct 23.
Key Takeaways
- ✓Scale: Super VL is 120B total / 12B active with up to 1M context; BF16 reference serving needs 4×H100 or 2×GB200, with an NVFP4 build for smaller footprints (model card)
- ✓Coding (vendor-reported): SWE-bench Verified 71.10, Terminal-Bench 2.1 66.70, LiveCodeBench v6 88.22; in the same table Qwen3.5-122B-A10B scores 71.9 / 48.70 / 76.7 and GLM-5.3-Flash 79.10 / 82.70 / 80.6
- ✓Agent evals: Hermes Bench 63.44% (Nemotron 3 Ultra 56.33%), LangChain Evals 87.05% (Ultra 84.3%); NVIDIA claims up to 5x single-GPU throughput vs open peers and 600+ TPS on Baseten (tech blog)
- ✓Bolt 74B-A9B 24GB: native 2-bit (IQ2 + QAT), 256K context; Terminal-Bench 2.1 50.98 and BrowseComp 66.75 vs 44.38 / 48.74 for Qwen3.6-35B-A3B; PinchBench 87.82 is slightly below Qwen3.6's 88.07 (Bolt card)
- ✓Serving: LiLiCorr speculative decoding gives up to 2.3x faster generation on DGX Spark; the Dynamo recipe serves 8x more users with 85% lower TTFT on 16×B200 at 50 tok/s/user
Not enough VRAM? Compare cloud API and self-hosting costs
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
Always-on enterprise agents read PDFs, charts and web pages and call several tools per task. Open stacks have usually chained a text LLM with a separate VLM, doubling latency and failure points, while closed frontier models cannot be post-trained on private infrastructure. NVIDIA's Oct 15 release packages document vision, tool use and open post-training into one model family, tiered for servers, desktops and edge devices.
Architecture
Nemotron 3.5 Super VL is a 120B-total / 12B-active Mamba-2–Transformer hybrid MoE with a C-RADIO vision encoder and up to 1M tokens of context. It was distilled from Nemotron 3 Ultra, trained on ~26T tokens, then post-trained on ~3.8M samples plus RL across dozens of environments. Inference features include multi-token prediction, the open LiLiCorr speculative decoder (up to 2.3x faster on DGX Spark) and NVFP4 with quantization-aware distillation. Bolt 74B-A9B 24GB uses native 2-bit QAT to fit 24GB for local agents on RTX/DGX Spark; Bolt E4B 4GB arrives Oct 23.
Benchmarks (vendor-reported)
Super VL: SWE-bench Verified 71.10, Terminal-Bench 2.1 66.70, BrowseComp 78, OSWorld-Verified 51.14, versus 71.9 / 48.70 / 58.1 / 51.5 for Qwen3.5-122B-A10B; GLM-5.3-Flash still leads on SWE-bench Verified at 79.10. Hermes Bench 63.44% and LangChain Evals 87.05% both beat Nemotron 3 Ultra. Bolt scores 50.98 on Terminal-Bench 2.1 vs 44.38 for Qwen3.6-35B-A3B. Independent third-party runs are not yet published.
Getting started
Try it via OpenRouter (a free 1M-context tier is listed) or NVIDIA NIM. Self-host with vLLM v0.32.0 using --tool-call-parser qwen3_coder --reasoning-parser nemotron_v3 and MTP speculative decoding; toggle reasoning with enable_thinking. For a single local box, start with Bolt (llama.cpp, Ollama, LM Studio, Unsloth, vLLM). Weights are under OpenMDW-1.1; customize with NeMo AutoModel/RL or ModelScope ms-swift.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.