Managed inference API
Nebius Token Factory
Token Factory serves open-weight models through shared OpenAI-compatible APIs and isolated dedicated endpoints. Dedicated deployments can select supported GPU configurations, regions and eligible custom weights.
Product facts checked October 2, 2026Editorial scores last reviewed September 3, 2026
Some outbound links use a first-party redirect hop for click counting. Commission is only claimed when a partner programme is active for that specific link — most vendor hops here are not paid placements. Affiliate disclosure.
Marketplace referrals (for example RunPod or Vast) may return credits or kickbacks when a programme is active — that is a material connection even when it is not a cash CPA.
Best for
Shared model APIs or configurable dedicated open-weight serving with region and capacity requirements
Watch out
Confirm exact checkpoints, regions and quotas in Token Factory. Nebius’s LPX announcement cites Artificial Analysis measuring 3,400 output tokens/sec for a single user on Gemma 4 31B; this specific model/hardware result does not guarantee throughput for other models, concurrent traffic or end-to-end latency.
| Billing | Public serverless per-token; dedicated deployments per GPU-hour with per-minute granularity. |
|---|---|
| GPU choice | Public shared capacity is platform-managed; dedicated deployments select supported GPU types and GPUs per replica. |
| Cold start | Dedicated capacity is governed by configured replicas and lifecycle; no universal shared or dedicated startup latency is established in the captured docs. |
| Model access | Shared endpoints serve base models. Dedicated endpoints use supported model/performance templates and can serve eligible custom/fine-tuned weights through the Custom Weights Hub. |
Sources for the dated product and announcement claims
- Current product modes, billing and deployment documentation — source checked October 2, 2026
- Current product modes, billing and deployment documentation — source checked October 2, 2026
- Nebius LPX announcement and benchmark attribution — source checked October 2, 2026
Fit detail
When Nebius Token Factory is the right shortlist — and when it is not
Use these lists to kill bad comparisons early, before you compare logos.
Ideal for
- Open-weight models through an OpenAI-compatible shared API
- Dedicated supported models with chosen region, GPU configuration and replica bounds
- Eligible fine-tuned weights with managed serving and observability
Usually not ideal for
- Raw GPU or custom CUDA access through Token Factory's managed serving API
- Arbitrary models outside supported templates and custom-weight eligibility
- Assuming public shared requests always remain in one chosen region
Spend & ops
How cost behaves — and what breaks first
No invented $/hour quotes. These notes explain the failure modes that show up on the first real invoice.
Cost mental model
Shared requests bill per token; dedicated endpoints bill GPU capacity by minute even though rates are shown per hour. Compare the exact template, replicas, region and utilization. A universal cheaper/premium provider ranking is not established.
Ops notes
- Public endpoints have dynamic rate limits and may change region with capacity.
- Dedicated endpoints reserve capacity, select a fixed region and scale within your replica bounds.
- Choose deployment mode before comparing token prices, GPU time or regional guarantees.
Fit scores
Nebius Token Factory on the decision axes
Same editorial scale as the hub chart — useful for shortlists, not as a price quote.
Nebius Token Factory
- GPU choice2/10
- Time-to-serving (editorial)4/10
- Price clarity4/10
- Production ops4/10
- Open-model breadth4/10
Among all 16 providers
GPU choice
Pick exact GPUs
Time-to-serving (editorial)
Warm-path fit
Price clarity
Easy to forecast
Production ops
Less DIY ops
Open-model breadth
Catalog depth
Scores are editorial planning ratings (1–10) for product shape — not published $/hour quotes or vendor SLAs. Verify current pricing on each provider site.
Related
Keep the decision attached to the rest of the stack
Cloud rental is one path. Local VRAM and open-weight fit still matter.