NVIDIA releases Nemotron 3.5 Super VL with 120B parameters and 1M-token context

• NVIDIA released a 120B-parameter multimodal model with 12B active parameters per token and up to 1M tokens of context, as described in the release overview.
• Factory and Ollama added day-one access: Factory reported stronger instruction-following and tool-call parallelism, while Ollama published the command ollama run nemotron-3.5-super.
• The throughput headline is up to 5x on a single GPU against similarly sized open models, according to Ollama’s announcement, which repeats NVIDIA’s claim.
• Customization arrived alongside inference: Fastino announced hosted SFT and RL support, and Pipecat’s release announcement introduced a Super VL fine-tune for voice agents.

The deployment recipes sample 1080p video at one frame per second and include a separate speculative decoder for DGX Spark. The neatest engineering detail is Mamba-state reuse: repeated agent calls preserve recurrent states alongside attention caches.

Released October 15, Super VL accepts text, images, and video and generates text, according to NVIDIA’s model card. Documents enter as page images with accompanying text; the listed video input is MP4, up to two minutes.

Its architecture has three main components:

• Language backbone: interleaved Mamba-2, attention, and mixture-of-experts layers, with 512 routed experts per MoE layer.
• Vision path: a C-RADIO encoder and connector that translate visual inputs into representations for the language backbone.
• Drafting head: native multi-token prediction for speculative decoding.

NVIDIA’s release includes base, post-trained, and NVFP4-quantized checkpoints. The model cards specify OpenMDW-1.1 licensing for commercial and non-commercial use.

Super VL turns a website screenshot into code and a rendered page

Super VL is available in Factory’s Droid. Factory described three observations from its initial use:

• Strong instruction-following, codebase exploration, and validation.
• Persistence on difficult tasks.
• Improved tool-call parallelism.

Ollama uses the shorter slug nemotron-3.5-super, omitting “VL,” in its launch command. Its model page also includes an Ollama-backed Claude Code launch command: ollama launch claude --model nemotron-3.5-super.

Super VL beat Nemotron 3 Ultra on two harness evaluations published in NVIDIA’s technical post:

| Evaluation | Nemotron 3 Ultra | Super VL | Change |
| --- | ---: | ---: | ---: |
| Hermes Bench | 56.33% | 63.44% | +7.11 points |
| LangChain Evals | 84.30% | 87.05% | +2.75 points |

The speed figures cover different workloads. In its launch post, vLLM reports more than 600 output tokens per second at 100,000 input tokens on the Baseten endpoint.

A separate chart reports 4x faster task completion for SWE-bench Pro, measuring the time to process 10,000 tasks on four B200 GPUs. The single-GPU, up-to-5x throughput claim also appears in Ollama’s announcement.

NVIDIA describes three mechanisms for reducing generation and repeated-prompt work in its inference breakdown:

• Multi-token prediction: an internal drafter proposes several future tokens, which the main model verifies in one forward pass. A dedicated training phase aligns the drafter with the final post-trained model.
• LiLiCorr: a separate speculative drafter evaluates compatibility across candidate tokens rather than predicting each position independently. NVIDIA reports up to 2.3x faster token generation on DGX Spark versus decoding without speculation.
• Context reuse: shared system instructions, tool definitions, and conversation prefixes reuse prior computation. Additional Mamba checkpoints reduce replay during speculative decoding, while internal checkpointing avoids extra model passes during prompt processing.

The NVFP4 card distinguishes the two drafting paths: native MTP needs no separate drafter download; LiLiCorr uses its own checkpoint.

The advertised context window is shared by the prompt, visual inputs, conversation history, tool results, and generated answer. NVIDIA’s model cards provide these hardware configurations:

| Checkpoint | Published configurations |
| --- | --- |
| BF16 | Four H100s or two GB200s |
| NVFP4 | One H200, two H100s, one GB200, or one DGX Spark |

The NVFP4 card lists 1M-token vLLM configurations for those platforms. Individual recipes can configure smaller windows to accommodate memory and concurrency.

Version and configuration details differ between the documentation surfaces. The vLLM launch post uses v0.31.0 and caps its example at 114,688 tokens; NVIDIA’s model cards specify v0.32.0 or a pinned nightly and include full-window recipes.

The cards also document the request controls:

• Tool parsing: qwen3_coder, with automatic tool choice enabled in the serving examples.
• Reasoning parsing: nemotron_v3 in vLLM.
• Reasoning levels: off, on (medium), and on; medium reduces generated tokens with a possible accuracy trade-off.
• Sampling: temperature 1.0 for text and 0.6 for vision, both with top_p=0.95.

The quantized checkpoint uses quantization-aware distillation to match the BF16 model. According to NVIDIA’s model card, Blackwell inference uses 4-bit weights and activations; non-Blackwell configurations use 4-bit weights with 16-bit activations.

NVIDIA holds harnesses, configurations, and other non-model parameters constant in its checkpoint comparison. Accuracy changes vary substantially by benchmark:

| Evaluation | BF16 | NVFP4 | Change |
| --- | ---: | ---: | ---: |
| SWE-bench Verified | 71.10 | 71.07 | −0.03 points |
| Terminal-Bench 2.1 | 60.67 | 62.78 | +2.11 points |
| MMLU-ProX | 76.91 | 73.76 | −3.15 points |
| WebVoyager, offline | 77.72 | 66.31 | −11.41 points |

The vision encoder, connector, Mamba convolutions, and embeddings remain in BF16; the MoE router remains in FP32.

Ollama also added a separate model, Nemotron 3.5 Bolt 74B-A9B, with native 2-bit quantization designed to fit in 24GB of memory alongside other applications.

Fastino’s Training API added day-zero Super VL support, with customer-controlled training loops, data, rewards, and evaluations. The service handles execution, sampling, gradients, weight updates, and checkpoints.

Fastino specifies three groups of primitives in its launch post:

• SFT: LoRA ranks 1 through 32, cross-entropy loss, AdamW, and gradient accumulation.
• RL: grouped sampling with token log probabilities, importance sampling, PPO, CISPO, and DRO.
• Evaluation and persistence: forward-only evaluation, seeded sampling, and checkpoints containing adapter and optimizer state.

Existing Tinker SDK workflows migrate by changing the server address and API key, Fastino says. Its Super VL rates per million tokens are:

• Prefill: $0.20, or $0.04 cached.
• Sampling: $0.72.
• Training: $0.96.

The model-specific table claims 25% to 65% savings against Tinker, narrower than the announcement’s headline 25% to 85% range. Checkpoint storage is billed separately.

Using an early-access checkpoint and text-only inputs, Fastino trained finance and healthcare adapters that improved all 15 held-out benchmarks it measured. The adapter updates 75.2M parameters, about 0.06% of Super VL, and the domain runs took two to 5.5 hours.

Pipecat’s PhoneLLM Max is a full-weights Super VL fine-tune for voice-agent pipelines. It emits one sentence of reasoning before answering, adding roughly 100ms to time to first answer token, according to the release thread.

On Pipecat’s own PhoneBench V1.1, Max with microthinking scores 83.0% versus 65.1% for base Super VL with thinking off, +17.9 points, in the launch chart. The comparison changes both the weights and the reasoning mode.

Alpha 2 also broadens prompting styles and task shapes and incorporates Pipecat Flows evaluations, as described in the release follow-up. Pipecat’s training account says the models learned from synthetic Flows sessions, with state-machine-guided conversations broken into smaller subtasks.

The advertised sub-200ms Max latency measures an individual inference call, according to a follow-up on production latency. Tool use requires two round trips, with network latency and inference-server queueing added; the stated production target is below 600ms P95 time to first answer token.