Nebius
Listed at https://nebius.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Per-token OpenAI-compatible inference: DeepSeek-V4-Flash $0.14/M input, $0.28/M output; Kimi K3 $3/$15; MiniMax M3 $0.30/$1.20; Llama 3.3 70B $0.13/$0.40 | per million tokens |
| Output | Two flavors per model - Base and Fast (-fast suffix) - same outputs, Fast trades higher token price for lower latency via speculative decoding | per million tokens |
🤖 Supported Models (29)
✨ Pros
- ✓AI-cloud spin-off from Yandex; NVIDIA holds ~9.3% stake, quarterly revenue up 684% YoY
- ✓Token Factory serves an OpenAI-compatible inference API across 29 open models incl. DeepSeek V4, Kimi K3, Qwen3, GLM-5.2
- ✓Same-account GPU cloud: scale from hosted inference to on-demand H100/H200/B200 GPUs and fine-tuning
- ✓Base/Fast dual flavors - Fast cuts latency with speculative decoding; dynamic rate limits auto-scale to 20x
- ✓Self-serve and instant: create account, get an API key, no quota approvals
⚠️ Cons
- ×No mainland China endpoint - proxy or aggregator required
- ×Catalog is third-party open weights only; no first-party frontier proprietary models (vs GPT-5, Claude)
- ×Token Factory docs are JS-rendered, English-only and developer-oriented
- ×Per-token price varies by model and GPU configuration (B200 vs H100); Base vs Fast spread needs evaluation
- ×Stronger brand in US/EU markets; fewer Chinese-language resources and community
🎯 Best For
International teams that want an OpenAI-compatible inference API plus same-account GPU compute and fine-tuning; developers who value an open-weights, vendor-neutral catalog; non-mainland users who can use a proxy
💰 Pricing & Plans
| Model | Input ($/M) | Output ($/M) | Context | Notes |
|---|---|---|---|---|
| DeepSeek-V4-Flash | $0.14 | $0.28 | 1024K | 1M-context reasoning, 292 tps (us-central1) |
| DeepSeek-V4-Pro | $1.75 | $3.50 | 1000K | Pro reasoning/coding tier (uk-south1) |
| Kimi-K3 | $3.00 | $15.00 | 1024K | Moonshot frontier open MoE (eu-west2) |
| MiniMax-M3 | $0.30 | $1.20 | 1049K | 428B MoE on B200, vLLM + EAGLE3 (us-central1) |
| Qwen3-235B-A22B-Instruct-2507 | $0.20 | $0.60 | 262K | Qwen flagship balanced chat (eu-north1) |
| GLM-5.2 | $1.40 | $4.40 | 1024K | Zhipu multimodal flagship (uk-south1) |
| Llama-3.3-70B-Instruct | $0.13 | $0.40 | 128K | Meta open-weights generalist (eu-north1) |
| Qwen3-Embedding-8B | $0.01 | $0.00 | 41K | Embedding; input-only billing (eu-north1) |
🔧 API & Developer Experience
- •API Surface: Token Factory exposes an OpenAI-compatible API at https://api.tokenfactory.nebius.com/v1/, so the standard OpenAI SDK works by swapping base_url and api_key. Text-to-text, vision (image2text) and embedding model types are all served through the same surface.
- •Auth & Signup: Create an account at tokenfactory.nebius.com with your Google or GitHub login, then generate an API key from the dashboard. No quota approval or sales call - the platform is fully self-serve for the startup tier.
- •Dual Flavors: Most text models offer Base and Fast flavors. Fast (append -fast to the model name) uses smaller batch sizes, more compute and speculative decoding to cut latency for the same output quality, at a higher per-token price.
- •Dynamic Rate Limits: Limits are elastic: measured in 15-minute rolling windows, they scale up 1.2x when average use stays at 80%+ of the cap and scale down otherwise, up to 20x the base allocation before an Enterprise plan is required.
- •Playground & Post-training: A UI playground lets you test and compare models. Fine-tuning (LoRA and full) and Data Lab dataset tooling are first-party features, and there are integrations with LangChain, LlamaIndex and the Hugging Face ecosystem.
🤖 Open-Weights Catalog & GPU Cloud
Nebius is the AI-cloud company spun out of Yandex and, as of July 2026, about 9.3% owned by NVIDIA - the stake disclosure sent its NASDAQ-listed stock (NBIS) up nearly 19% in a single session, and its Q1 revenue jumped roughly 684% year over year on AI data-center demand. Token Factory is its hosted inference product, offering 29 open-weights models from DeepSeek V4, Kimi K3, MiniMax M3, Qwen3, GLM-5.2 and the NVIDIA Nemotron family, plus embeddings. The same account can move into raw GPU capacity and fine-tuning, making Token Factory the API gateway to Nebius's broader GPU cloud - a rare combination of a neutral open-model marketplace and direct hardware access under one roof. Its positioning line is, in its own words, 'build and scale faster on the purpose-built AI cloud, engineered from silicon to API.'
🌐 Regional Availability & Latency
Token Factory serves models from regional GPU fleets that today span North America and Europe - us-central1 (US), eu-north1 and eu-west2 (EU) and uk-south1 (UK) are the regions active in the public model catalog, with different flagship models pinned to specific locations (for example DeepSeek-V4-Flash in us-central1, Kimi-K3 in eu-west2, GLM-5.2 in uk-south1). There is no mainland China regional endpoint and no official China access program, so production use from China requires a stable overseas proxy or an aggregator that fronts the Token Factory API. The OpenAI-compatible surface keeps client-side integration latency low; for latency-critical serving each model also exposes a Fast flavor plus dedicated endpoints that can be pinned to a chosen region for more predictable performance.