What was released

On August 11, 2026, NVIDIA published the open weights for Nemotron 3.5 Lightning 30B-A3B. It is a text reasoning and chat model with a hybrid architecture combining Mamba-2, mixture-of-experts, and attention layers. The model contains 30 billion parameters in total and activates 3 billion for each token.

On the same day, Ollama 0.32.9 added the Nemotron 3 architecture and a ready-to-run `nemotron-3.5-lightning` entry. Ollama lists several variants: `latest` is a 25 GB package with a context window up to one million tokens, while the MLX variant is 23 GB with a 256K context window. The files are available for download; this is not a placeholder or countdown announcement.

For a small or medium business, the important point is not another leaderboard. This release represents a practical class of local model: per-token computation is closer to three billion active parameters, while quality and specialization draw on a much larger total set of weights. That trade-off can suit always-on agents, task routing, RAG, and automation—provided infrastructure is sized for the complete artifact rather than the attractive A3B label.

Three billion active is not a three-billion-parameter model

An MoE model selects a subset of experts for every token. This reduces computation compared with a dense 30B model, but it does not remove the need to store and load all experts. Active parameters primarily describe per-token computation; total parameters drive storage, memory, loading, and image-distribution requirements.

The official BF16 card specifies one H100 80 GB or A100 80 GB for single-GPU deployment. NVIDIA recommends the NVFP4 variant for optimized inference, while Ollama offers quantized packages of roughly 23–25 GB. The weights are only part of the budget: KV cache, runtime workspaces, context, concurrent requests, and operating-system headroom must be added.

This creates two management implications:

  • “3B active” does not mean the model fits into 4–8 GB of memory;
  • “up to 1M context” does not mean a million tokens is economical on the selected machine.

Long context increases KV-cache demand and prompt-processing time. For an enterprise assistant, good document chunking and retrieval of a few relevant passages is usually cheaper than placing an entire archive into every prompt.

What is confirmed and what still needs validation

According to the developer card, the model was trained on more than 20 trillion tokens and supports configurable reasoning, structured outputs, and tool use. NVIDIA publishes its own reasoning, coding, and agentic results together with reproducibility recipes. Treat those numbers as vendor claims until they are independently reproduced in the target workflow.

The supported natural-language list includes English, Spanish, French, German, Italian, and Japanese. Russian is not listed. That does not prove the model cannot answer in Russian, but it makes a dedicated Russian evaluation mandatory: company terminology, entity inflection, extraction of business identifiers, refusal when evidence is missing, and tool calls with Cyrillic arguments.

The BF16 weights are intended mainly for post-training, distillation, research, and creation of custom quantizations. The card recommends NVFP4 for direct serving. Ollama lowers the barrier to a first run, but a one-command installation does not replace quality testing of the exact quantization and runtime.

License: commercial use is allowed, but this is not Apache 2.0

The model uses OpenMDW-1.1. The card describes it as ready for commercial use. The license permits using, modifying, and distributing model materials without charge, while requiring a copy of the agreement and applicable notices to remain with distributions.

It also terminates granted rights in certain voluntary patent or copyright infringement lawsuits. Users remain responsible for third-party rights and required permissions. This will not usually block an internal pilot, but legal staff should record the license, the source revision, and rules for passing a derivative model to a contractor or customer.

Do not apply the code repository's license to the weights by assumption. Different project artifacts can use different terms. A model registry should store the license and revision of the exact checkpoint.

Where the model can help a business

A first reasonable use is a local agent for internal operations where predictable cost and privacy matter more than universal knowledge. Examples include:

  • classifying and routing incoming requests;
  • RAG over internal procedures with source citations;
  • drafting answers for an employee;
  • extracting structured fields from already recognized text;
  • calling read-only tools for order, inventory, or ticket status;
  • reviewing exception queues and preparing a report.

A second use is specialized adaptation. Full BF16 weights and published recipes provide a base for SFT, RL, or distillation when a company has high-quality labelled data and a clear metric. Post-training should not be the default first step: RAG, a strict schema, and strong prompt examples often solve the task at lower cost.

A poor first use is granting an “always-on agent” immediate access to email, CRM, file systems, and payments. A model designed for agents is not a guarantee of safe autonomy.

Build a safe local agent stack

The model server should generate a proposed action only. Execution passes through a separate tool gateway that checks user identity, role, parameters, and the policy for the specific operation.

A minimum design includes:

1. a local model endpoint with no direct internet exposure;
2. a RAG service that filters documents by user permissions;
3. an orchestrator limiting steps, wall time, and token budget;
4. a tool gateway with operation allowlists and strict JSON schemas;
5. mandatory human approval for writes, sends, and deletes;
6. an audit trail of prompts, retrieved sources, proposed calls, human decisions, and results;
7. a kill switch that revokes agent access without stopping unrelated systems.

Reading an order status can be one narrowly scoped method with a single identifier. Changing a price or sending an email needs a separate route and approval. If the model sends an extra field or an out-of-range value, the gateway rejects the call instead of guessing its intent.

Evaluate it in five working days

Day 1: choose one process and collect 100–200 privacy-treated examples. Separate common requests, edge cases, and cases where the system must refuse.

Day 2: pin two serving options—for example, Ollama `latest` and one optimized server format. Record download size, startup time, peak memory, and every component version.

Day 3: measure Russian and business-domain quality. For RAG, score citation correctness; for extraction, field exact match; for an agent, tool-name and argument validity.

Day 4: load-test realistic prompt lengths. Capture p50 and p95 latency, tokens per second, concurrency, memory, and errors. Test long contexts separately: the card maximum should not become the default setting.

Day 5: run in shadow mode. The model proposes an answer or action, while employees continue the current process. Scale only after predefined quality, time, and cost gates pass.

A modeled economic case

Assume an internal assistant handles 12,000 requests per month. An employee spends four minutes on search and drafting. A pilot reduces that to two minutes for 60% of requests and does not change the rest.

The affected 7,200 requests previously consumed 480 hours and now consume 240 hours, releasing 240 hours—not 480. At a fully loaded labor cost of RUB 1,100 per hour, the modeled resource effect is RUB 264,000 per month.

If servers, electricity, support, and quality control cost RUB 120,000 per month, the modeled net effect is RUB 144,000. With RUB 1.15 million in implementation cost, simple payback is about eight months.

This is an example, not a promise. Replace the assumptions with the observed useful-answer rate, review cost, equipment utilization, and error cost. Released time becomes financial value only when it reduces overtime, response time, or future hiring.

What managers should take away

Nemotron 3.5 Lightning is interesting because it combines 30 billion total with 3 billion active parameters, downloadable weights, and Ollama support. But “3B active” is not a memory estimate, and a model that does not list Russian as supported should not reach employees without an internal evaluation.

The practical next step is not purchasing a server. Run a five-day test of one process on an existing staging machine. Pin a Russian golden set, two configurations, a context limit, tool policies, and a modeled economic case. If the model passes the gates, then choose permanent infrastructure and a weight format.