Nebius：模型智能、性能与价格

Nebius

本分析旨在帮助你根据使用场景，选择 Nebius 提供的最佳模型。

最智能

Updated

Kimi K3.1 Pro

DeepSeek V4.1 Flash 1017 (max)

GLM-5.3 (max) (FP4)

MiniMax-M3.1-Flash

GLM-5.3-Flash

Intelligence Index

共 27 个模型

每任务耗时最短

Beam 1.0 (NVFP4)

1.9 min

Qwen3.8 27B (non-reasoning) (FP4)

2.4 min

Nemotron 3 Ultra

4.3 min

gpt-oss-120b (high) (Base)

5.8 min

Nemotron 3.5 Lightning (BF16)

6.2 min

每个 Intelligence Index 任务耗时（分钟）

共 27 个模型

价格最低

Nemotron 3.5 Lightning (BF16)

$0.08

Nemotron 3 Nano

$0.08

MiniMax-M3.1-Flash

$0.10

GLM-5.3-Flash

$0.10

Gemma 3 27B (FP8)

$0.12

每 100 万 token 的混合价格

共 27 个模型

Nebius 提供 27 个模型，每个模型的智能、性能和价格特征各不相同。下方对比了各模型的关键指标。

智能方面，Nebius 上表现最好的模型是 Kimi K3.1 Pro（46）、DeepSeek V4.1 Flash 1017 (max)（33）和GLM-5.3 (max) (FP4)（32）。
每任务耗时方面，Beam 1.0 (NVFP4)（1.9 分钟）、Qwen3.8 27B (non-reasoning) (FP4)（2.4 分钟）和Nemotron 3 Ultra（4.3 分钟）在 Nebius 上完成一个 Intelligence Index 任务最快。各模型之间的每任务耗时差异显著，最慢的模型耗时是最快模型的 3.3 倍。
延迟方面，gpt-oss-120b (low) Base（1.35 秒）、GLM-5.2 (non-reasoning) (FP4)（1.56 秒）和Qwen3.8 27B (non-reasoning) (FP4)（1.66 秒）的首个答案 Token 延迟最低。
价格方面，Nemotron 3.5 Lightning (BF16)（$0.08）、Nemotron 3 Nano（$0.08）和MiniMax-M3.1-Flash（$0.10）每 100 万 token 的混合价格最低。
上下文窗口方面，MiniMax-M3 (FP8)（1M）、Kimi K3.1 Pro（1M）和DeepSeek V4.1 Flash 1017 (max)（1M）支持 Nebius 上最大的上下文窗口。

Updated

智能

Artificial Analysis Intelligence Index · Higher is better

每任务耗时

Inference time (minutes) per Intelligence Index task · Lower is better

价格

USD per 1M tokens (blended) · Lower is better

智能Updated

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v5.0.1 incorporates 9 evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1

Artificial Analysis Intelligence Index v5.0.1 includes: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), Terminal-Bench-Science 0.1. See Intelligence Index methodology for further details, including a breakdown of each evaluation and how we run them.

Evaluations

Evaluations measured independently by Artificial Analysis · AA-PersonalityBench scores are percentile ranks from 1 to 100

AA-Briefcase v1.1

Agentic knowledge work, Elo

AA-Omniscience v1.1 AccuracyUpdated

Knowledge

AA-Omniscience v1.1 Non-Hallucination RateUpdated

1 - hallucination rate

AutomationBench-AA

Agentic SaaS workflows

Terminal-Bench-Science 0.1New

Agentic scientific research workflows in a terminal

FrontierCode-AA 2.0New

Agentic software engineering in open-source repositories

HLE-DiamondNew

Reasoning & knowledge

GDP.pdf

Professional document reasoning, All-pass

Terminal-Bench 4.0

Agentic coding & terminal use

CritPt v2 (Hard Subset)

Physics reasoning

AA-AnalystAgent

Quantitative analysis on spreadsheets & documents

Harvey LAB-AA v1.1Updated

Legal agentic work, Hallucination-Gated All-Pass Rate

EnterpriseOps-Gym-AA

Agentic business operations

GDPval-AA v2.1

Agentic real-world work tasks, Elo

AA-3DNew

Agentic interactive 3D experiences, Elo

Code2Video-AANew

Product and brand videos designed in code, Elo

ITBench-AA

Kubernetes incident root-cause analysis

MLCR-AAUpdated

Medical long context reasoning

MMMU-Pro

Visual reasoning

AA-PersonalityBench: Expressiveness ScoreNew

Rich, varied, well-crafted language

AA-PersonalityBench: Organization ScoreNew

Well-structured, scannable, answer-first

AA-PersonalityBench: Transparency ScoreNew

Truthful, acknowledges uncertainty, non-sycophantic

AA-PersonalityBench: Slop ScoreNew

Use of filler and generic language

While model intelligence generally translates across use cases, specific evaluations may be more relevant for certain use cases.

Intelligence Index 与价格

Blended at 7:2:1 (cache-input-output) · USD per 1M tokens (blended)

Most attractive quadrant

Pareto line

While higher intelligence models are typically more expensive, they do not all follow the same price-quality curve.

价格

每项 Intelligence Index 任务的成本

Weighted average cost (USD) per Intelligence Index task · Segmented by token type · Lower is better

Weighted average cost per Intelligence Index task. Each evaluation’s cost is calculated from input, cache hit, cache write, reasoning, and answer token prices, divided by task count, and weighted by its Intelligence Index weight.

Intelligence Index 与价格

Blended at 7:2:1 (cache-input-output) · USD per 1M tokens (blended)

Most attractive quadrant

Pareto line

While higher intelligence models are typically more expensive, they do not all follow the same price-quality curve.

速度

已测量的单次请求和多轮智能体工作负载的平均值

Intelligence Index 每项任务耗时

Output tokens divided by end-to-end output speed (minutes per Intelligence Index task) · Lower is better

The weighted average time (minutes) per Artificial Analysis Intelligence Index task: output tokens per task divided by output speed, weighted by each benchmark's weight in the Intelligence Index.

Speed and latency metrics average each endpoint's runs over the past 5 days, and the error rate pools them. We run the benchmark every 6 hours, and endpoints of deprecated models every 18 hours.

质量

Endpoint Accuracy Index

各接入点准确率相对其模型基准接入点的百分比 · 颜色表示差距是否具有统计显著性

Composite measure of how much of a model's accuracy a given provider endpoint preserves, from re-running BFCL v4-500, HLE-250 and AA-LCR-25 against that endpoint. Where a self-hosted reference endpoint exists, scores are expressed as a percentage of that reference (100 = matches reference); lower scores indicate accuracy lost to quantisation, sampling defaults, or other endpoint-side configuration. Scores are point-in-time snapshots. Methodology.

延迟

已测量的单次请求和多轮智能体工作负载的平均值

首个答案 Token 延迟

经过推理后首个答案 token 到达所需的秒数 · 已测量的单次请求和多轮智能体工作负载的平均值 · 越低越好

Time to first answer token received, in seconds, after API request sent. For reasoning models, this includes the 'thinking' time of the model before providing an answer. For models which do not support streaming, this represents time to receive the completion.

Speed and latency metrics average each endpoint's runs over the past 5 days, and the error rate pools them. We run the benchmark every 6 hours, and endpoints of deprecated models every 18 hours.

可靠性

在所有已测量的单次请求和多轮智能体工作负载中，接入点失败的请求占比

接入点错误率

接入点因错误、限速或超时而失败的请求占比 · 涵盖所有已测量的单次请求和多轮智能体工作负载 · 越低越好

Share of requests over the past 5 days that the endpoint failed with an error, a rate limit or a timeout. Failures on our side are left out of both counts. Across several workloads, the failures and requests are pooled before dividing.

Speed and latency metrics average each endpoint's runs over the past 5 days, and the error rate pools them. We run the benchmark every 6 hours, and endpoints of deprecated models every 18 hours.

缓存

缓存命中率

多轮智能体会话各次运行在最近四周的中位数，更新于 2026年10月26日 · 越高越好

The share of cacheable input tokens served from the provider's prompt cache, measured during multi-turn agentic sessions. We regularly measure each endpoint and report a representative rate based on recent measurements. Higher values mean more of the input is billed at the cheaper cache-hit price, where the provider offers one, which lowers the effective cost of agentic workloads. For further details, see our methodology page.

每任务成本与缓存命中率

Weighted average cost (USD) per Intelligence Index task · Cache hit rate: median of the last four weeks, updated Oct 26, 2026

Most attractive quadrant

Pareto line


								进一步分析
Kimi K3.1 Pro	46	$4.38	11.8	65.51	2.4%	1M	开放	模型服务商
DeepSeek V4.1 Flash 1017 (max)	33	$3.27	20.7	64.78	0.0%	1M	开放	模型服务商
GLM-5.3 (max) (FP4)	32	$4.92	17.9	41.62	0.9%	1M	开放	模型服务商
MiniMax-M3.1-Flash	29	$1.01	13.7	31.53	0.0%	1M	开放	模型服务商
GLM-5.3-Flash	28	$0.95	17.1	55.70	0.0%	1M	开放	模型服务商
Muse Spark Open (xhigh) (FP4)	25	$1.61	8.7	14.79	1.5%	1M	开放	模型服务商
DeepSeek V4 Pro 0813 (max)	23	$13.43	18.7	48.41	0.8%	1M	开放	模型服务商
Qwen3.8 27B (xhigh) (FP4)	20	$1.98	6.2	20.13	1.8%	262k	开放	模型服务商
MiniMax-M3 (FP8)	19	$1.30	7.5	5.22	0.0%	1M	开放	模型服务商
Beam 1.0 (NVFP4)	18	$0.65	1.9	5.70	0.0%	1M	开放	模型服务商
Kimi K3 (low)	17	$1.81	9.2	67.78	13.5%	1M	开放	模型服务商
Nemotron 3.5 Super VL	14	$0.84	9.7	17.48	3.1%	1M	开放	模型服务商
DeepSeek V4 Pro 0813 (non-reasoning)	13	$6.52	15.6	2.94	0.0%	1M	开放	模型服务商
Nemotron 3 Ultra	12	$3.81	4.3	4.68	0.0%	256k	开放	模型服务商
Qwen3.8 27B (non-reasoning) (FP4)	10	$2.33	2.4	1.66	0.0%	262k	开放	模型服务商
Nemotron 3.5 Lightning (BF16)	6	$0.32	6.2	21.39	0.0%	1M	开放	模型服务商
gpt-oss-120b (high) (Base)	4	$0.35	5.8	13.45	8.7%	128k	开放	模型服务商
Gemma 3 27B (FP8)	--	--	--	17.85	0.0%	110k	开放	模型服务商
Nemotron 3 Super	--	--	--	7.69	2.6%	256k	开放	模型服务商
Nemotron 3 Nano	--	--	--	6.52	1.4%	262k	开放	模型服务商
Qwen3 30B A3B 2507 (non-reasoning)	--	--	--	2.95	0.0%	262k	开放	模型服务商
Hermes 4 405B (FP8)	--	--	--	18.78	2.9%	128k	开放	模型服务商
gpt-oss-120b (low) Base	--	--	--	1.35	7.4%	128k	开放	模型服务商
Hermes 4 405B (non-reasoning) (FP8)	--	--	--	5.16	1.7%	128k	开放	模型服务商
GLM-5.2 (non-reasoning) (FP4)	--	--	--	1.56	0.0%	432k	开放	模型服务商
GLM-5.2 (max) (FP4)	--	--	--	20.80	0.0%	432k	开放	模型服务商
Kimi K3 (max)	--	--	--	75.39	3.4%	1M	开放	模型服务商

关键定义

Average cost per task in the index. Costs are split by input, cache hit, cache write, reasoning, and answer token pricing where canonical token counts are available.

常见问题

关于 Nebius 的常见问题

Nebius 上最智能的模型是 Kimi K3.1 Pro，Intelligence Index 得分为 46。

Nebius 上每任务耗时最短的模型是 Beam 1.0 (NVFP4)，每个 Intelligence Index 任务耗时 1.9 分钟。每任务耗时等于每任务输出 token 数除以端到端输出速度，数值越低表示任务完成得越快。

Nebius 上首个答案 Token 延迟最低的模型是 gpt-oss-120b (low) Base，延迟为 1.35 秒。延迟越低，初始响应越快。

按混合价格计算，Nebius 上最实惠的模型是 Nemotron 3.5 Lightning (BF16)，每 100 万 token 的价格为 $0.078（缓存命中/输入/输出比例为 7:2:1）。

Nebius 上各模型价格最多相差 29.6 倍，从最实惠的 Nemotron 3.5 Lightning (BF16)（每 100 万 token $0.078）到最昂贵的 Kimi K3 (max)（每 100 万 token $2.31）。

是，Nebius 提供兼容 OpenAI 的 API，可以轻松从 OpenAI 切换，或继续使用现有的 OpenAI SDK 集成。

是，Nebius 上全部 27 个模型均支持用于结构化输出的 JSON 模式。

是，Nebius 上全部 27 个模型均支持函数调用（工具使用）。

是，Nebius 提供 21 个推理模型：Kimi K3.1 Pro、DeepSeek V4.1 Flash 1017 (max)、GLM-5.3 (max) (FP4)、MiniMax-M3.1-Flash、GLM-5.3-Flash、Muse Spark Open (xhigh) (FP4)、DeepSeek V4 Pro 0813 (max)、Qwen3.8 27B (xhigh) (FP4)、MiniMax-M3 (FP8)、Beam 1.0 (NVFP4)、Kimi K3 (low)、Nemotron 3.5 Super VL、Nemotron 3 Ultra、Nemotron 3.5 Lightning (BF16)、gpt-oss-120b (high) (Base)、Nemotron 3 Super、Nemotron 3 Nano、Hermes 4 405B (FP8)、gpt-oss-120b (low) Base、GLM-5.2 (max) (FP4)和Kimi K3 (max)。推理模型会先进行扩展思考来解决复杂问题，再给出答案。

是，Nebius 上全部 27 个模型均为开放权重模型。

会。服务商性能可能因基础设施变化、负载均衡和更新而随时间波动。我们持续对所有服务商进行基准测试，并在“随时间变化”图表中展示历史性能趋势。

选择 Nebius 上的模型时，请考虑：智能（适合质量敏感型任务）、每任务耗时（完成一个 Intelligence Index 任务的速度）、延迟（适合需要快速首次响应的交互式应用）、价格（适合成本敏感型工作负载），以及上下文窗口大小、JSON 模式或函数调用支持等功能。

Nebius：模型智能、性能与价格

最智能

每任务耗时最短

价格最低

智能

每任务耗时

价格

智能Updated

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index

Evaluations

Intelligence Evaluation Relevance

Artificial Analysis Intelligence Index

Intelligence Index 与价格

Price-Quality Variance

价格

每项 Intelligence Index 任务的成本

Cost per Intelligence Index Task

Intelligence Index 与价格

Price-Quality Variance

速度

Intelligence Index 每项任务耗时

Time per Intelligence Index Task

Time Period

质量

Endpoint Accuracy Index

Endpoint Accuracy Index

延迟

首个答案 Token 延迟

Time to First Answer Token

Time Period

可靠性

接入点错误率

Endpoint Error Rate

Time Period

缓存

缓存命中率

What Cache Hit Rate Shows

每任务成本与缓存命中率

Cost per Intelligence Index Task

What Cache Hit Rate Shows

关键定义

Index Cost per Task

常见问题

Nebius 提供哪些模型？

Nebius 上最智能的模型是什么？

Nebius 上哪个模型的每任务耗时最短？

Nebius 上哪个模型的延迟最低？

Nebius 上最便宜的模型是什么？

Nebius 上各模型之间的价格差异有多大？

Nebius 是否兼容 OpenAI？

Nebius 是否支持 JSON 模式？

Nebius 是否支持函数调用？

Nebius 是否提供推理模型？

Nebius 是否提供开放权重模型？

Nebius 的性能会随时间变化吗？

如何选择 Nebius 上的模型？