Use · Nebius (Token Factory)

Is it good enough?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 64.4/100

This page is a projection of the one entry record, the Reliability factor that Use covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Models served

The catalogue (tokenfactory.nebius.com/models/catalog, rendered in a browser) lists 133 models, including DeepSeek V4.1 Flash and V4.1 Flash 1017, DeepSeek-V4-Pro-0813 and DeepSeek-R1-0528; Kimi-K3, Kimi-K3.1-Pro and Kimi-K2.5/K2.6/K2.7-Code; GLM-5.1 to GLM-5.3, GLM-5.3-Flash and MiniMax-M3; Qwen3.x families including Qwen3.8-27B and Qwen3-Coder-480B-A35B; gpt-oss-120b and gpt-oss-20b, Muse Spark Open and Reflection Beam 1.0; NVIDIA Nemotron-3-Ultra, Nemotron 3.5 Super VL and 3.5-Lightning; and Llama 3.x, Gemma 4, Hermes-4 and embedding models.

The platform column shows the same tags on every row, so the per-model page decides whether a public endpoint exists. It does for gpt-oss-120b, DeepSeek-V4-Pro-0813, GLM-5.3 and MiniMax-M3. DeepSeek-V4-Pro, Llama-3.3-70B-Instruct and Kimi-K2-Instruct are dedicated-only. The first two follow Serverless removal notices dated 2026-11-09 and 2026-08-31. The serverless catalogue changes every few weeks, so check the model page before building on it.

Inference features

Endpoints are OpenAI-compatible chat/completions. Per-model tool-calling and structured-output support follows each served model's own capabilities and is unverified on Nebius' endpoints, so confirm per model.

Fine-tuning / custom models

A fine-tuning feature is documented (supervised fine-tuning, custom speculator training and a data lab). The Legal Quick Guide states fine-tuning artifacts are stored exclusively in EU data centres with no regional selection. Supported methods and base models are not stated.

Context & token limits

The catalogue lists a context window per model, for example 131K tokens for gpt-oss-120b, 979K for DeepSeek-V4-Pro-0813, 1,024K for GLM-5.3 and Kimi-K3 and 128K for Llama-3.3-70B-Instruct. About a fifth of the rows show no value.

How this scores

The ownership factor this domain covers, drawn from the one entry record.

3

ReliabilityDoes it stay up and stay secure?

Moderate

No inference SLA exists, and Supplemental Terms 7.4 says so expressly (no service level commitments, uptime guarantees or credits unless agreed in writing). Self-service dedicated endpoints carry no formal SLA and release capacity on any stop or incident. The status page shows a Token Factory component per region and two resolved Token Factory incidents in its 14-day window.

How this scores (AOI sub-dimensions)
Reliability3/5whether it stays up, with an SLA and status historyToken Factory has no inference SLA, and Supplemental Terms 7.4 (effective 2026-10-22) excludes one: Nebius "does not provide any service level commitments, uptime guarantees, or associated compensation for the Token Factory Service".
Security3/5the controls protecting your traffic and dataThe trust centre documents customer workload isolation (VPC, InfiniBand and Kubernetes), TLS in transit, storage encryption by default (with named exceptions), least-privilege access with MFA, vulnerability management and regular security assessments, a structured incident-response plan and a secure SDLC.
Compliance5/5which independent certifications and attestations it holdsTwo independent artefacts support a 5.
What this means for adoptionData control is limited by default. The binding Token Factory Supplemental Terms (5.1(b), effective 2026-10-22) allow Nebius to train its smaller Speculative Decoding models on your Inputs and Outputs unless you opt out. The Legal Quick Guide tells readers that content is not used to train any model. Compliance is a strength, evidenced by a public Deloitte SOC 3 report and an SII-QCD ISO 27001 certificate. Two EU data-plane regions (Finland, France) are documented, but public endpoints are unpinned and the standard terms carry no inference SLA. For sensitive data, enable Zero Data Retention (it also stops Speculative Decoding) and use a dedicated endpoint in an EU region for residency.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Terms of serviceread2026-11-19
Nebius Token Factory Supplemental Terms (publication and effective date 2026-10-22), incorporated into the Nebius Services Agreement.
Terms of serviceread2026-11-19
Legacy Nebius Token Factory Terms of Service (publication date 2026-08-20).
Documentationread2026-11-19
Nebius Legal Quick Guide (non-binding; states that the Terms, DPA and Privacy Policy prevail).
Terms of serviceread2026-11-19
Nebius Services Agreement (effective 2026-10-22).
Data Processing Addendumread2026-11-19
Nebius Data Processing Addendum (effective 2026-11-18).
Subprocessorsread2026-11-19
Nebius Token Factory sub-processor list (effective 2026-09-23): 23 table rows and 20 distinct entities in three groups.
Securityread2026-11-19
Nebius Trust Center (static pages) lists SOC 2 Type II with HIPAA (report under NDA; auditor Deloitte; scope includes AI Cloud, AI Studio (now Token Factory) and Tracto AI), SOC 3 (public).
Third-party analysisread2026-11-19
Public SOC 3 report for Nebius BV, 'Security, Availability and Confidentiality + HIPAA Mapping', period April 01, 2025 to July 31, 2025, independent service auditor's report dated October 19, 2025 (Deloitte, per the trust-centre page and the report footer).
Third-party analysisread2026-11-19
ISO/IEC 27001:2022 certificate no.
Documentationread2026-11-19
Nebius AI Cloud regions documentation.
Slaread2026-11-19
Nebius SLA pages.
Slaread2026-11-19
Nebius public status page (status.nebius.com): All Systems Operational.
Documentationread2026-11-19
Nebius Token Factory model catalogue and per-model pages (rendered in a browser without login; curl gets HTTP 403 from a Cloudflare challenge).
Documentationread2026-11-19
Nebius Token Factory Model Deprecations notices.
Documentationread2026-11-19
Nebius Token Factory docs.
Documentationread2026-11-19
Nebius dedicated-endpoint billing-policy docs: 'A dedicated endpoint is considered active, accessible, and billable when at least one replica is running'; no charge when zero replicas run; 'Charges may vary depending on your custom contract or work order'.
Documentationread2026-11-19
Nebius Token Factory product page and quickstart: serverless inference 'through a familiar API' and dedicated endpoints 'on isolated capacity' with 'regional placement'.