Nebius Token Factory API 2026: The Open-Weights GPU Cloud Gateway

Nebius is the AI-cloud company spun out of Yandex and, as of July 2026, roughly 9.3% owned by NVIDIA — a stake disclosure that sent its NASDAQ stock (NBIS) up nearly 19% in a single session. Its Q1 revenue jumped about 684% year over year on AI data-center demand. Nebius Token Factory is the hosted-inference product that turns that GPU-cloud muscle into an OpenAI-compatible API: 29 open-weights models from DeepSeek V4, Kimi K3, MiniMax M3, Qwen3, GLM-5.2 and NVIDIA's Nemotron family, plus embeddings, with a Base and Fast flavor per model. This review covers the API, verified August 2026 pricing, dual flavors, dynamic rate limits, the broader GPU cloud, regional access and the competitive picture.

🤖 Model catalog

Token Factory's catalog is entirely open weights, which makes it deliberately vendor-neutral: one API key routes to DeepSeek V4, Moonshot's Kimi K3, MiniMax M3, Alibaba's Qwen3, Zhipu's GLM-5.2 and Meta's Llama 3.3 alongside NVIDIA's Nemotron 3 family and Nous Research's Hermes-4. Model types span text-to-text, vision (image2text) and embeddings. There is no first-party frontier proprietary model, so if you specifically want GPT-5, Claude or Gemini you would combine Token Factory with another provider — see our OpenAI-compatible APIs guide for how to route between them.

For a cheaper-focused comparison of open-weights serving, our DeepInfra, Together AI and Groq reviews cover the fast-and-cheap end; the cheapest LLM APIs roundup ranks them all.

💰 Pricing (verified 2026-08-12)

Token Factory is pay-as-you-go per token — no setup fee, no monthly minimum. Prices below are pulled from Nebius's own public model catalog via its machine-readable catalog, which is the authoritative source.

The low end is where Token Factory competes hardest: DeepSeek-V4-Flash at $0.14/$0.28 and Qwen3-235B at $0.20/$0.60 land in the same budget tier as the cheapest open-weights hosts. Each model also has a Fast flavor (append -fast to the model id) that raises per-token price in exchange for lower latency via speculative decoding — a useful lever for interactive apps. For a strategy view of managing these per-token costs across providers, see our AI API cost control guide.

🔌 OpenAI-compatible inference API

from openai import OpenAI

client = OpenAI(
    base_url="https://api.tokenfactory.nebius.com/v1/",
    api_key="YOUR_NEBIUS_API_KEY"
)
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4-Flash",
    messages=[{"role": "user", "content": "Summarize Nebius in one sentence."}]
)
print(response.choices[0].message.content)

Because the surface is OpenAI-compatible, any tool that speaks the OpenAI SDK — LangChain, LlamaIndex, AutoGen, CrewAI, the Vercel AI SDK — works by swapping base_url and api_key. The official docs cover the endpoint, model list and rate limits. Sign-up is fully self-serve: create an account with Google or GitHub at tokenfactory.nebius.com, generate an API key, and start calling. For a Python-native serverless alternative, compare Modal; for another GPU-cloud-plus-API hybrid launched this month, see our Hyperbolic review.

🚀 Getting started

The fastest path is to sign up at tokenfactory.nebius.com with your Google or GitHub account, copy an API key from the dashboard, and point the standard OpenAI client at https://api.tokenfactory.nebius.com/v1/ using the snippet above. Because rate limits auto-scale and there is no credit card required to inspect the catalog, you can spend the first few hours comparing DeepSeek-V4-Flash against Qwen3-235B or GLM-5.2 on your own prompts in the playground at near-zero cost. When a workload stabilizes, evaluate the Fast flavor for interactive latency, then decide whether to move fine-tuning or raw GPU work into the same Nebius account as your usage grows.

⚙️ Base vs Fast flavors & dynamic rate limits

Two details make Token Factory's DX distinctive. First, dual flavors: for most text models the Base flavor optimizes for cost and the Fast flavor (append -fast) uses smaller batches, more compute and speculative decoding to cut latency while keeping identical outputs — a clean way to pay only for the responsiveness you actually need. Second, dynamic rate limits: the platform measures usage in 15-minute rolling windows and auto-scales your cap up 1.2x when you sustain 80%+ utilization and back it down otherwise, up to 20x your base allocation before an Enterprise plan is required. In practice most teams on the startup tier never think about limits, which is a meaningful difference from fixed-RPM providers. Fine-tuning (LoRA and full) and Data Lab dataset tooling are also first-party, so a prototype can graduate to a tuned model without leaving the platform.

🖥️ The GPU cloud behind the API

Token Factory is the API gateway to Nebius's broader GPU cloud: the same account that calls the hosted endpoint can provision on-demand H100/H200/B200 GPU capacity for training, fine-tuning and batch work. Nebius describes itself as "the purpose-built AI cloud, engineered from silicon to API," and the 684% revenue surge reflects demand for that infrastructure. NVIDIA's ~9.3% stake — disclosed July 2026 — is both a capital and an ecosystem signal: Nebius sits on the supply side of NVIDIA's kit while also pricing open-weights inference competitively. That vertical integration is the core structural difference from pure inference hosts like Together AI or Fireworks, and closer to RunPod or Hyperbolic, which pair an OpenAI-compatible layer with raw GPU access.

🆚 vs DeepInfra / Together AI / Groq / Hyperbolic

The closest comparisons are the open-weights hosts. DeepInfra and Together AI are pure inference-first platforms with very broad catalogs and aggressive pricing; Groq owns the extremely-low-latency niche on its LPU hardware. Nebius differentiates on two axes: a same-account upgrade path into raw GPU and fine-tuning, and a catalog that is genuinely neutral — DeepSeek, Moonshot, Zhipu and Qwen models side by side with NVIDIA's own Nemotron line. Hyperbolic is the closest GPU-cloud hybrid, but Nebius adds fine-tuning and a bigger first-party infra narrative. If you primarily need Chinese mainland access, our DeepSeek and Qwen/Bailian reviews cover providers with mainland endpoints.

🌐 Regional access & latency

Token Factory serves models from GPU fleets across North America and Europe. The public catalog's active regions are us-central1 (US), eu-north1 and eu-west2 (EU) and uk-south1 (UK), with different flagships pinned to specific locations — for example DeepSeek-V4-Flash in us-central1, Kimi-K3 in eu-west2 and GLM-5.2 in uk-south1. There is no mainland China regional endpoint and no official China access program, so production use from China needs a stable overseas proxy or an aggregator that fronts the API. For latency-critical work, the Fast flavor plus dedicated endpoints (which can be pinned to a chosen region) give more predictable serving than shared multi-tenant scheduling.

🎯 Who is Nebius Token Factory for?

Three groups fit best. Teams that want vendor-neutral open models get DeepSeek, Qwen, GLM, Kimi and Nemotron under one OpenAI-compatible key — useful when you want to swap weights without changing providers. Startups that will outgrow proxy inference can start on the hosted API and later move fine-tuning or raw GPU work into the same Nebius account rather than porting to a separate cloud. Latency-conscious builders benefit from the Base/Fast split and dynamic rate limits that remove fixed-RPM ceilings. The trade-off is that it is not a frontier-proprietary-model vendor, and there is no mainland China path.

⚠️ Limits to know

• No mainland China endpoint — production access requires a proxy or aggregator.
• Open weights only: no first-party frontier proprietary models (compare GPT-5, Claude, Gemini).
• Per-token price varies by the GPU configuration behind each model (B200 vs H100), and the Base/Fast spread must be evaluated for your workload.
• Docs and dashboard are developer-oriented, English-only.
• Brand is stronger in US/EU markets; fewer Chinese-language resources and community.
• Rate-limit auto-scaling is dynamic — plan for the 20x elastic ceiling rather than assuming a fixed RPM.
• Dedicated endpoints and raw GPU capacity are priced separately from hosted inference.

🎯 Verdict

Choose Nebius Token Factory when you want a genuinely neutral open-weights catalog at budget prices and the ability to grow into GPU compute and fine-tuning in the same account. DeepSeek-V4-Flash at $0.14/$0.28 and Qwen3-235B at $0.20/$0.60 put it squarely in the cheapest tier, while NVIDIA's stake and the 684% revenue growth point to serious infrastructure behind it. The Fast flavor and 20x dynamic rate limits make it friendly to latency-sensitive and spiky workloads. For pure lowest-cost/broadest-catalog open inference, DeepInfra and Together AI remain strong; for a GPU-cloud hybrid with raw-GPU marketplaces, compare Hyperbolic and RunPod. Nebius is the differentiated pick for teams that want open models and direct hardware under one roof.

🔗 Official resources

• Nebius Token Factory (app, model catalog, API keys)
• Token Factory Documentation (quickstart, endpoint, rate limits)
• Nebius — purpose-built AI cloud (company homepage)

FAQ

Is Nebius Token Factory OpenAI compatible?
Yes. Point the OpenAI client base_url at https://api.tokenfactory.nebius.com/v1/ and use standard chat-completions calls.

How much does Nebius Token Factory cost?
Pay-as-you-go per token: DeepSeek-V4-Flash $0.14/$0.28, Qwen3-235B $0.20/$0.60, MiniMax-M3 $0.30/$1.20, Llama-3.3-70B $0.13/$0.40, GLM-5.2 $1.40/$4.40, Kimi-K3 $3/$15 per million. No setup fee or monthly minimum.

Is there a free tier?
No permanent free tier, but no setup fee and no monthly subscription — you pay only for tokens used, with dynamic rate limits that auto-scale to 20x.

What is the Fast flavor?
Append -fast to a model id to trade a higher per-token price for lower latency via smaller batches and speculative decoding, with identical output quality.

Can it be used from China?
Not directly — it needs a proxy or aggregator; no mainland regional endpoint exists (active regions are us-central1, eu-north1, eu-west2, uk-south1).

Does Nebius offer GPU and fine-tuning?
Yes — the same account can provision on-demand H100/H200/B200 GPUs and run LoRA or full fine-tuning plus Data Lab dataset tooling.