Vercel puts NVIDIA Nemotron 3.5 Lightning on AI Gateway for free

What did Vercel announce?

Vercel's changelog entry Nemotron 3.5 Lightning now available on AI Gateway for free, dated 14 October 2026 and credited to Zachary Chen and Jerilyn Zheng, says the model is live as nvidia/nemotron-3.5-lightning-free and is served by NVIDIA. The piece is short and product-shaped: it names the slug, sketches the model, shows copy-paste snippets for the AI SDK and OpenAI-compatible Chat Completions, and points coding-agent users at npx vercel ai-gateway setup.

That matters because AI Gateway already sits in many teams' paths for Claude Code, Codex, Cursor and similar agents. Adding a free NVIDIA open-weight endpoint means those agents can pick a hosted Nemotron without a separate NVIDIA key or a self-hosted GPU box, as long as they accept Vercel's Gateway auth and NVIDIA's prompt-retention warning.

The changelog frames Nemotron as an open-weight reasoning and chat model for long-running agents, sub-agent work, chatbots, and RAG systems. It does not claim a new model release on 14 October. The dated object is distribution: free access through Vercel's unified API, with NVIDIA as the serving provider.

What is Nemotron 3.5 Lightning?

NVIDIA's model card for NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 on Hugging Face lists 30 billion total parameters with 3 billion active per token, a hybrid architecture that interleaves Mamba-2 and MoE layers with selected attention layers, and a context length of up to 1 million tokens. The card's release date is 11 August 2026; the license is OpenMDW License Agreement version 1.1. Recommended sampling is temperature 1.0 and top_p 0.95.

The same card says the model is aimed at long-running autonomous agents, sub-agent workhorse deployments, and efficient local inference on personal hardware. Supported languages include English plus coding languages, Spanish, French, German, Italian and Japanese. Hardware notes cover Blackwell (including DGX Spark / GB10 and GB200), Hopper H100/H200, and Ampere via W4A16, with speculative-decoding options such as DSpark, MTP and DFlash for some setups.

That lineage sits beside earlier Nemotron work AI Lookout has covered, including NVIDIA's Nemotron IMO/IOI results and on-device boxes such as DGX Spark with 64 GB. The Vercel post is not a new training announcement; it is a free hosted path for a model that already had open weights and NVIDIA-hosted options elsewhere.

Vercel's own summary matches the card on the 30B / 3B active split, the hybrid MoE design, the 1M context window, and tool calling plus structured outputs. Teams evaluating the model should still read NVIDIA's card for precision formats, speculative-decoding recipes and license obligations rather than treating the changelog as a full model card.

How do you call the free model on AI Gateway?

For application code, Vercel shows an AI SDK streamText call with model set to nvidia/nemotron-3.5-lightning-free, and a Python OpenAI client pointed at https://ai-gateway.vercel.sh/v1 with AI_GATEWAY_API_KEY. The changelog also walks Claude Code and fx users through AI Gateway setup: run npx vercel ai-gateway setup, then pick Nemotron 3.5 Lightning from /model. Setup is said to turn on gateway model discovery so the picker lists the full catalog.

That pattern is useful for teams that already standardize on one Gateway key across agents. Instead of teaching every tool a different NVIDIA base URL, they keep Vercel's endpoint and change only the model string. Failover, budgets and usage reporting stay in the Gateway control plane Vercel describes in its AI Gateway docs.

Vercel states that AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including on Bring Your Own Key (BYOK) requests. For this particular slug the listed price is free, so the practical cost is whatever quotas, rate limits or prompt-logging terms NVIDIA attaches to the free served endpoint, not a Vercel per-token fee.

What does free mean in practice?

Free here is a Gateway catalog claim, not a promise that NVIDIA will host unlimited traffic forever. The changelog's explicit caveat is that the provider may retain prompts for model improvement. That is the main product trade-off versus self-hosting the OpenMDW weights or paying for a private endpoint where prompts stay inside your contract.

For prototypes, internal coding agents and RAG experiments that already route through Vercel, the free slug removes a billing step and a second provider account. For regulated workloads, customer data, or prompts you cannot send to a provider that may retain them, the free path is the wrong tool even if the model quality is fine.

Open-weight local stacks remain an alternative when privacy or offline use matters. Projects such as Hermes Agent's local-model connectors and efficient on-device models like Liquid AI's LFM2.5-24B-A2B target a different constraint: keep inference on hardware you control. Vercel's offer is complementary, hosted convenience with an explicit retention warning, not a replacement for those setups.

How this fits open-weight agent stacks

Agent builders have been assembling stacks from open weights, gateways and MCP-style tools. A free long-context workhorse on a popular gateway lowers the cost of spinning up sub-agents for research, coding or RAG while keeping orchestration in tools developers already use.

The model's 1M-token context and tool-calling support match that use case on paper. Whether it beats a paid Claude or GPT workhorse on a given task is an empirical question Vercel does not answer in the changelog. The dated fact is availability and price point, not a new independent leaderboard result.

Operators who self-host should also keep security hygiene in mind. Serving stacks such as vLLM have seen a burst of advisories this month; see our note on eighteen vLLM security advisories in early October. Choosing the free Gateway path avoids that ops burden while introducing provider-side prompt retention instead.

Limits, privacy, and what is still unclear

Vercel does not publish per-key free-tier quotas, rate limits or a sunset date in the changelog entry we reviewed. Teams that plan production traffic on the free slug need to confirm current limits in the Gateway dashboard and NVIDIA's serving terms, not assume unlimited capacity from the word free.

The entry also does not say whether BYOK NVIDIA credentials can attach to the same model under different retention rules, or how the free slug interacts with enterprise data-processing agreements. Those gaps matter as soon as a pilot leaves a sandbox.

We have not independently load-tested nvidia/nemotron-3.5-lightning-free through AI Gateway. Numbers for size, context and architecture above come from Vercel's changelog and NVIDIA's Hugging Face model card as read on 15 October 2026. Treat vendor performance claims as vendor claims until you measure them on your own tasks.