Compare models
Put up to 4 models beside each other — token prices, context windows, capabilities and provider, from the same catalogue the model pages read.
| Attribute | Nemotron 3.5 Super VLnemotron-3.5-super-vl | DeepSeek V4.1 Flash 1017deepseek-v4.1-flash-1017 | DeepSeek V3.2 Specialedeepseek-v3.2-speciale |
|---|---|---|---|
| Pricing | |||
| Input | $0.10 / 1M | $0.10 / 1M | $0.28 / 1M |
| Output | $0.40 / 1M | $0.50 / 1M | $0.40 / 1M |
| Cache Write (5m) | $0.10 / 1M | $0.10 / 1M | $0.28 / 1M |
| Cache Write (1h) | $0.10 / 1M | $0.10 / 1M | $0.28 / 1M |
| Cache Read | $0.10 / 1M | $0.10 / 1M | $0.28 / 1M |
| Web Search | $0 / 1M | $0 / 1M | $0 / 1M |
| Context | |||
| Max context | 1M | 1M | 163.8K |
| Max output | N/A | N/A | N/A |
| Capabilities | |||
| Vision | Yes | Yes | Yes |
| Function Calling | Yes | Yes | Yes |
| JSON Mode | Yes | Yes | Yes |
| Streaming | Yes | Yes | Yes |
| Catalogue | |||
| Provider | NVIDIA | DeepSeek | DeepSeek |
| Category | chat | chat | chat |
| Charge type | Pay As You Go | Pay As You Go | Pay As You Go |
| Released | — | — | — |
| Description | |||
| Summary | NVIDIA Nemotron 3.5 Super VL is an open multimodal Mixture-of-Experts (MoE) model featuring 120B total parameters with 12B activated per token. Built on a hybrid Mamba–Transformer architecture with multi-token prediction, it supports text, images, documents, GUI screenshots, and video inputs, along with a 1M-token context window for complex, long-horizon workflows. Optimized for multimodal reasoning, visual understanding, and agentic tool use, the model supports configurable reasoning modes and token budgets. It is well suited for document intelligence, web research, coding validation, customer service, GUI analysis, and video understanding, enabling efficient, long-running multimodal agent applications. | DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek’s cost-efficient Mixture-of-Experts (MoE) model, built on its innovative Causal Encoder-Decoder (CED) architecture. Featuring 552B total parameters with 8B activated for input and 16B for output, it combines efficient inference with native multimodal understanding through jointly trained text and visual representations. Optimized for coding, terminal and computer-use agents, and long-horizon workflows, the model introduces compressed KV caching that reduces cache memory usage to approximately one-quarter of the previous Flash generation. DeepSeek reports that it surpasses V4 Pro in performance, inference speed, and task completion time, making it well suited for cost-efficient autonomous agents, complex software engineering, and long-context analysis. | DeepSeek-V3.2-Speciale is a high-compute edition of V3.2 built for top-tier reasoning and agent performance. Using DeepSeek Sparse Attention and extensive reinforcement learning, it surpasses GPT-5 on tough reasoning benchmarks and approaches Gemini 3 Pro–level capability, while still remaining strong at coding and tool use. It also draws on a large agent-training pipeline to boost reliability and generalization in interactive environments. |