Qwen3.8 27B Competitive Intelligence ReportQwen3.8 27B 全模型竞争情报报告

A complete audit of all 610 models on Artificial Analysis Intelligence Index v4.1.1, organised around Qwen3.8 27B (II = 52.02) as the benchmark anchor. 围绕 Qwen3.8 27B（II = 52.02）作为基准锚点，对 Artificial Analysis Intelligence Index v4.1.1 上全部 610 个模型进行系统审计。

Source: artificialanalysis.ai · Generated: 2026-08 · Methodology: 9-benchmark composite II (GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR) · Costs shown where publicly disclosed; N/A means no public price 数据来源: artificialanalysis.ai · 生成时间: 2026-08 · 评测方法: 9 项基准综合（GDPval-AA v2、τ³-Banking、Terminal-Bench v2.1、SciCode、Humanity's Last Exam、GPQA Diamond、CritPt、AA-Omniscience、AA-LCR）· 价格仅显示公开报价；"暂无"表示未公开

Anchor Model基准模型

52.02Intelligence IndexIntelligence Index

Qwen3.8 27B · Alibaba, China · 27B params · Released 2026-08-14阿里巴巴 (中国) · 27B 参数 · 发布 2026-08-14

Public prices: input $0.45 / 1M tok, output $3.20 / 1M tok, cache hit $0.05 / 1M tok. No public per-task cost. 公开计费（按 1M tokens）：输入 $0.45、输出 $3.20、缓存命中 $0.05。未公开每任务成本。

Output tokens / task: 47,166 (reasoning 37,214 + answer 9,952) 每任务输出 tokens：47,166（其中推理 37,214 + 答案 9,952）

Reasoning model · Open weights · OPEN 推理型模型 · 开源权重 · 开源

Position in the Index总分定位

Total models模型总数	610
With II score含 II 分数	597
Unique model families独立模型族	421
Rank排名	~ top 25全球前 25
Quantile分位数	~ 96th percentile约 96 百分位
Open-weight peers强过它的开源	only 5仅 5 open models are stronger个开源模型
Closed peers强过它的闭源	14 closed共 14 are stronger个闭源模型

Verdict结论

Open-weight frontier at 27B.27B 级别的开源前沿。 Qwen3.8 27B is the highest-scoring 27B-class dense model (open or closed) and ties or beats every Anthropic Sonnet, every GPT-5.x pre-Luna, every GPT-5.2 / 5.3 / 5.4, and ties GPT-5.6 Luna (the OpenAI "small reasoning" tier) to within 0.30 II. Qwen3.8 27B 是当前 27B 稠密模型（开源/闭源）最高分。它与所有 Anthropic Sonnet、所有 GPT-5.x pre-Luna、GPT-5.2 / 5.3 / 5.4 持平或领先，并把 OpenAI 的"小型推理"档 GPT-5.6 Luna 压到仅 0.30 II 之外。

Headline:头条： Anthropic Opus 5 (63.05), Claude Fable 5 (62.07), GPT-5.6 Sol (60.93), and Grok 4.6 (60.92) are the only 4 models more than 8 II points ahead. 仅 Claude Opus 5（63.05）、Claude Fable 5（62.07）、GPT-5.6 Sol（60.93）、Grok 4.6（60.92）这 4 个模型在 II 上领先 Qwen3.8 27B 超过 8 分。

Distribution of the 420 unique families420 个独立模型族的分布

Grouping by base name (stripping reasoning-level suffixes like (max), (xhigh), (high), (medium), (low), (Non-reasoning), and date suffixes), and reporting the strongest variant's II. Qwen3.8 27B itself is excluded from the totals. 按基础名分组（去除推理档位后缀如 (max)、(xhigh)、(high)、(medium)、(low)、(Non-reasoning) 与日期后缀），取每族最强变体的 II。Qwen3.8 27B 本身不计入总数。

Stronger强于 Qwen

II > 54.02

2 open开源 · 14 closed闭源

Similar实力相当

II ∈ [50.02, 54.02]

3 open开源 · 5 closed闭源

Weaker弱于 Qwen

396

II < 50.02

Of top 60:前 60 中： 37 open开源 · 23 closed闭源

Competitive position on the II scale (0 – 70)在 II 标尺（0 – 70）上的竞争位置

Qwen 52

Qwen3.8 27B

Stronger open更强开源

Stronger closed更强闭源

Similar (± 2 II)相当 (± 2 II)

Stronger强于 16 models above Qwen3.8 27B (II > 54.02) 16 个模型超过 Qwen3.8 27B (II > 54.02)

Anthropic dominates with 5 entries. Only 2 open-weight models surpass Qwen3.8 27B: Kimi K3 and Qwen3.8 2.4T A95B. Anthropic 一家占据 5 席。仅 2 个开源模型能稳压 Qwen3.8 27B：Kimi K3 与 Qwen3.8 2.4T A95B。

II	Δ	Name名称	Creator厂商	Country国家	Params参数	Type类型	Cost成本
63.05	+11.03	Claude Opus 5 high, low, medium, xhigh	Anthropic	US	large	closedR	$2.3369
62.07	+10.05	Claude Fable 5 —	Anthropic	US	large	closedR	$3.1396
60.93	+8.91	GPT-5.6 Sol Non-reasoning, high, low, medium	OpenAI	US	large	closedR	$1.2312
60.92	+8.90	Grok 4.6 —	SpaceXAI	US	large	closedR	$0.8367
59.70	+7.68	Kimi K3 low	Kimi	CN	2800B	openR	$0.8375
59.51	+7.49	GLM-5.3 —	Z AI	CN	753B	closedR	$0.6829
58.08	+6.06	Qwen3.8 Max —	Alibaba	CN	large	closedR	$1.1320
57.70	+5.68	Qwen3.8 2.4T A95B —	Alibaba	CN	2400B	openR	$1.0919
57.33	+5.31	Claude Opus 4.8 —	Anthropic	US	large	closedR	$2.0315
56.76	+4.74	Muse Spark 1.2 —	Meta	US	large	closedR	$0.3992
56.58	+4.56	GPT-5.6 Terra Non-reasoning, high, low, medium	OpenAI	US	large	closedR	$0.5080
56.31	+4.29	GPT-5.5 Non-reasoning, high, low, medium	OpenAI	US	large	closedR	$1.1747
56.03	+4.01	Gemini 3.7 Flash low, medium	Google	US	large	closedR	$0.4022
55.76	+3.74	Grok 4.5 —	SpaceXAI	US	large	closedR	$0.3601
55.26	+3.24	Claude Sonnet 5 Non-reasoning	Anthropic	US	large	closedR	$1.7173
54.96	+2.94	Claude Opus 4.7 Non-reasoning, high	Anthropic	US	large	closedR	$2.2285
63.05	+11.03	Claude Opus 5 high, low, medium, xhigh	Anthropic	美国	large	闭源推理	$2.3369
62.07	+10.05	Claude Fable 5 —	Anthropic	美国	large	闭源推理	$3.1396
60.93	+8.91	GPT-5.6 Sol Non-reasoning, high, low, medium	OpenAI	美国	large	闭源推理	$1.2312
60.92	+8.90	Grok 4.6 —	SpaceXAI	美国	large	闭源推理	$0.8367
59.70	+7.68	Kimi K3 low	Kimi	中国	2800B	开源推理	$0.8375
59.51	+7.49	GLM-5.3 —	Z AI	中国	753B	闭源推理	$0.6829
58.08	+6.06	Qwen3.8 Max —	Alibaba	中国	large	闭源推理	$1.1320
57.70	+5.68	Qwen3.8 2.4T A95B —	Alibaba	中国	2400B	开源推理	$1.0919
57.33	+5.31	Claude Opus 4.8 —	Anthropic	美国	large	闭源推理	$2.0315
56.76	+4.74	Muse Spark 1.2 —	Meta	美国	large	闭源推理	$0.3992
56.58	+4.56	GPT-5.6 Terra Non-reasoning, high, low, medium	OpenAI	美国	large	闭源推理	$0.5080
56.31	+4.29	GPT-5.5 Non-reasoning, high, low, medium	OpenAI	美国	large	闭源推理	$1.1747
56.03	+4.01	Gemini 3.7 Flash low, medium	Google	美国	large	闭源推理	$0.4022
55.76	+3.74	Grok 4.5 —	SpaceXAI	美国	large	闭源推理	$0.3601
55.26	+3.24	Claude Sonnet 5 Non-reasoning	Anthropic	美国	large	闭源推理	$1.7173
54.96	+2.94	Claude Opus 4.7 Non-reasoning, high	Anthropic	美国	large	闭源推理	$2.2285

Highlights — stronger强者要点

Claude Opus 5 (63.05, $2.34)Claude Opus 5（63.05，$2.34） is the current SOTA, +11.03 ahead of Qwen3.8 27B. Anthropic holds the top 2 spots (Opus 5 and Fable 5). 是当前 SOTA，比 Qwen3.8 27B 高 +11.03。Anthropic 拿下前 2 席（Opus 5 与 Fable 5）。

Open-weight ceiling:开源天花板： Kimi K3 (59.70) and Qwen3.8 2.4T A95B (57.70) are the only opens above Qwen3.8 27B. The 2.4T-A95B is the same Qwen3.8 family but in MoE 2.4T-total / 95B-active form — the obvious "next step up" within Qwen family. Kimi K3（59.70）与 Qwen3.8 2.4T A95B（57.70）是仅有的两个在 Qwen3.8 27B 之上的开源。2.4T-A95B 属于同一 Qwen3.8 家族，但采用 MoE 2.4T 总参数 / 95B 激活——是 Qwen 家族向上发展的"下一步"。

OpenAI's 4 modelsOpenAI 4 个模型 above Qwen3.8 27B: GPT-5.6 Sol (60.93), GPT-5.6 Terra (56.58), GPT-5.5 (56.31). All reasoning tier, ranging $0.51 to $1.23 / task. 在 Qwen3.8 27B 之上：GPT-5.6 Sol（60.93）、GPT-5.6 Terra（56.58）、GPT-5.5（56.31）。均为推理档，定价 $0.51 – $1.23 / 任务。

Similar相当 8 models in the ± 2 II band (II in [50.02, 54.02]) 8 个模型在 ± 2 II 区间 (II ∈ [50.02, 54.02])

The "dead heat" zone. Qwen3.8 27B sits between GPT-5.6 Luna (above) and Gemini 3.5 Flash (below). "贴身肉搏"区。Qwen3.8 27B 夹在 GPT-5.6 Luna（强侧）与 Gemini 3.5 Flash（弱侧）之间。

II	Δ	Name名称	Creator厂商	Country国家	Params参数	Type类型	Cost成本
52.02	± 0	Qwen3.8 27B the anchor model	Alibaba	CHINA	27B	openR	N/A
53.20	+1.18	DeepSeek V4 Pro 0813 —	DeepSeek	CN	1600B	openR	$0.2521
53.20	+1.18	Muse Spark 1.1 —	Meta	US	large	closedR	$0.2923
53.12	+1.10	GPT-5.4 Non-reasoning, low	OpenAI	US	large	closedR	$1.1038
52.64	+0.62	GLM-5.2 —	Z AI	CN	753B	openR	$0.4445
52.32	+0.30	GPT-5.6 Luna Non-reasoning, high, low, medium	OpenAI	US	medium	closedR	$0.0471
51.96	-0.06	Gemini 3.5 Flash medium, minimal	Google	US	large	closedR	$0.6934
51.77	-0.25	DeepSeek V4 Flash 0731 —	DeepSeek	CN	284B	openR	$0.1122
51.58	-0.44	Gemini 3.6 Flash —	Google	US	large	closedR	$0.3442
52.02	± 0	Qwen3.8 27B 基准锚点	阿里巴巴	中国	27B	开源推理	暂无
53.20	+1.18	DeepSeek V4 Pro 0813 —	DeepSeek	中国	1600B	开源推理	$0.2521
53.20	+1.18	Muse Spark 1.1 —	Meta	美国	large	闭源推理	$0.2923
53.12	+1.10	GPT-5.4 Non-reasoning, low	OpenAI	美国	large	闭源推理	$1.1038
52.64	+0.62	GLM-5.2 —	Z AI	中国	753B	开源推理	$0.4445
52.32	+0.30	GPT-5.6 Luna Non-reasoning, high, low, medium	OpenAI	美国	medium	闭源推理	$0.0471
51.96	-0.06	Gemini 3.5 Flash medium, minimal	Google	美国	large	闭源推理	$0.6934
51.77	-0.25	DeepSeek V4 Flash 0731 —	DeepSeek	中国	284B	开源推理	$0.1122
51.58	-0.44	Gemini 3.6 Flash —	Google	美国	large	闭源推理	$0.3442

Highlights — similar相当区间要点

4 opens vs 4 closeds该区间 4 开源 vs 4 闭源 in the ± 2 II band. DeepSeek V4 Pro 0813 (53.20) and Meta Muse Spark 1.1 (53.20) tie for the top of this band. 。DeepSeek V4 Pro 0813（53.20）与 Meta Muse Spark 1.1（53.20）并列首位。

OpenAI's closest equal: GPT-5.6 Luna (52.32, $0.05).OpenAI 最近的对等：GPT-5.6 Luna（52.32，$0.05）。 This is OpenAI's "small reasoning" tier — almost identical cost profile to Qwen3.8 27B per task, just 0.30 II ahead. 这是 OpenAI 的"小型推理"档——成本结构与 Qwen3.8 27B 几乎一致，仅高 0.30 II。

Google's closest equal: Gemini 3.5 Flash (51.96, $0.69).Google 最近的对等：Gemini 3.5 Flash（51.96，$0.69）。 Just 0.06 II below Qwen3.8 27B — within margin of error. 仅比 Qwen3.8 27B 低 0.06 II——误差范围内。

DeepSeek V4 Pro 0813 (53.20, $0.25)DeepSeek V4 Pro 0813（53.20，$0.25） is the highest-scoring open-weight model with public pricing that's directly comparable to Qwen3.8 27B. 是公开定价下与 Qwen3.8 27B 最直接可比、得分最高的开源模型。

Weaker弱于 Top 60 weaker models — the "obvious beats" (best variant shown) 前 60 个弱于 Qwen 的模型 — "明显落败"（每族最强变体）

The 396 weaker models are split into a sample of the top 60 here. Notice that even top-tier closed models like Claude Opus 4.6, GPT-5.2, and Claude Sonnet 4.6 (max) fall below Qwen3.8 27B. 396 个弱于 Qwen3.8 27B 的模型，这里展示前 60 个。值得注意的是，即便是顶级闭源模型 Claude Opus 4.6、GPT-5.2、Claude Sonnet 4.6 (max) 都低于 Qwen3.8 27B。

Premium weaker (II 35 – 50)高端弱于（II 35 – 50）

II	Δ	Name名称	Creator厂商	Country国家	Params参数	Type类型	Cost成本
48.37	-3.65	Claude Sonnet 4.6 Non-reasoning, Non-reasoning, Low Effort	Anthropic	US	large	closedR	$1.2182
47.74	-4.28	Gemini 3.1 Pro Preview —	Google	US	large	closedR	$0.3346
47.36	-4.66	Motif 3 Beta	Motif Technologies	KR	314B	c/oR	N/A
46.71	-5.31	Qwen3.7 Max —	Alibaba	CN	large	closedR	$0.5413
45.51	-6.51	GPT-5.3 Codex —	OpenAI	US	large	closedR	N/A
45.40	-6.62	MiniMax-M3 —	MiniMax	CN	428B	openR	$0.1387
45.27	-6.75	DeepSeek V4 Pro , high	DeepSeek	CN	1600B	openR	$0.0474
45.14	-6.88	Kimi K2.6 —	Kimi	CN	1000B	openR	$0.3653
44.93	-7.09	Claude Opus 4.6 high	Anthropic	US	large	closedR	N/A
44.25	-7.77	Muse Spark —	Meta	US	—	closedR	N/A
43.34	-8.68	GPT-5.2 , medium	OpenAI	US	large	closedR	N/A
43.02	-9.00	Kimi K2.7 Code —	Kimi	CN	1000B	openR	$0.2220
42.88	-9.14	MiMo-V2.5-Pro —	Xiaomi	CN	1023B	openR	$0.0339
42.29	-9.73	Inkling —	Thinking Machines	US	975B	openR	$0.3389
42.21	-9.81	Hy3 —	Tencent	CN	299B	openR	$0.0357
42.12	-9.90	DeepSeek V4 Flash , high	DeepSeek	CN	284B	openR	$0.0673
41.87	-10.15	Claude Opus 4.5 —	Anthropic	US	large	closedR	N/A
41.74	-10.28	Nex-N2-Pro —	Nex AGI	CN	397B	openR	N/A
41.64	-10.38	Solar Pro 4 —	Upstage	KR	medium	closedR	$0.2231
41.37	-10.65	MiMo-V2-Pro —	Xiaomi	CN	—	closedR	N/A
41.22	-10.80	GPT-5.2 Codex —	OpenAI	US	large	closedR	N/A
41.18	-10.84	Inkling Small —	Thinking Machines	US	266B	openR	$0.0735
41.07	-10.95	Qwen3.6 Max Preview —	Alibaba	CN	large	closedR	N/A
40.97	-11.05	GLM-5.1 —	Z AI	CN	744B	openR	$0.3042
40.94	-11.08	GPT-5.4 mini , medium	OpenAI	US	large	closedR	$0.4948
40.65	-11.37	Grok Build 0.1 0616 —	SpaceXAI	US	large	closedR	$0.2252
40.61	-11.41	Gemini 3 Pro Preview low	Google	US	large	closedR	N/A
40.55	-11.47	GLM-5 —	Z AI	CN	744B	openR	N/A
40.49	-11.53	Qwen3.6 Plus —	Alibaba	CN	large	closedR	$0.3569
39.85	-12.17	JT-4.1 Flash 236B A21B —	China Mobile	CN	236B	closed	N/A
39.71	-12.31	GPT-5.4 nano —	OpenAI	US	medium	closedR	$0.1493
39.71	-12.31	Agnes 2.5 Pro Alpha —	Sapiens AI	SG	medium	closedR	N/A
39.37	-12.65	Qwen3.7 Plus —	Alibaba	CN	medium	closedR	$0.2424
39.05	-12.97	GLM-5-Turbo —	Z AI	CN	—	closedR	N/A
38.87	-13.15	MiniMax-M2.7 —	MiniMax	CN	230B	openR	$0.0777
38.74	-13.28	Gemini 3 Flash —	Google	US	large	closedR	N/A
38.32	-13.70	Nemotron 3 Ultra —	NVIDIA	US	550B	openR	$0.3827
38.04	-13.98	MiMo-V2.5 —	Xiaomi	CN	310B	openR	$0.0104
37.95	-14.07	Grok 4.3 Non-reasoning, low, medium	SpaceXAI	US	large	closedR	$0.1454
37.95	-14.07	Grok 4.20 0309 v2 —	SpaceXAI	US	large	closedR	N/A
37.82	-14.20	Ling 3.0 Flash —	InclusionAI	CN	124B	openR	$0.0377
37.70	-14.32	Qwen3.6 27B —	Alibaba	CN	27.8B	openR	$0.2918
37.47	-14.55	GPT-5.1 —	OpenAI	US	large	closedR	$0.3040
37.44	-14.58	Grok 4.20 0309 —	SpaceXAI	US	large	closedR	N/A
37.44	-14.58	Gemini 3.5 Flash-Lite —	Google	US	medium	closedR	$0.0965
37.43	-14.59	Solar Open2 250B —	Upstage	KR	250B	openR	N/A
37.39	-14.63	Claude 4.5 Sonnet —	Anthropic	US	large	closedR	$0.4643
37.32	-14.70	MiMo-V2-Omni-0327 —	Xiaomi	CN	—	closedR	N/A
37.04	-14.98	GPT-5 Codex —	OpenAI	US	large	closedR	N/A
36.02	-16.00	Kimi K2.5 —	Kimi	CN	1000B	openR	$0.0964
35.87	-16.15	MiMo-V2-Omni —	Xiaomi	CN	—	closedR	N/A
35.60	-16.42	GPT-5.1 Codex —	OpenAI	US	large	closedR	N/A
35.35	-16.67	GLM 5V Turbo —	Z AI	CN	—	closedR	N/A
35.31	-16.71	GPT-5 ChatGPT, low, medium, minimal	OpenAI	US	large	closedR	$0.2572
35.06	-16.96	Muse Glimmer —	Meta	US	30B	openR	$0.0732
35.01	-17.01	A.X-K2 —	SK Telecom	KR	692B	openR	N/A
48.37	-3.65	Claude Sonnet 4.6 Non-reasoning, Non-reasoning, Low Effort	Anthropic	美国	large	闭源推理	$1.2182
47.74	-4.28	Gemini 3.1 Pro Preview —	Google	美国	large	闭源推理	$0.3346
47.36	-4.66	Motif 3 Beta	Motif Technologies	韩国	314B	双推理	N/A
46.71	-5.31	Qwen3.7 Max —	Alibaba	中国	large	闭源推理	$0.5413
45.51	-6.51	GPT-5.3 Codex —	OpenAI	美国	large	闭源推理	N/A
45.40	-6.62	MiniMax-M3 —	MiniMax	中国	428B	开源推理	$0.1387
45.27	-6.75	DeepSeek V4 Pro , high	DeepSeek	中国	1600B	开源推理	$0.0474
45.14	-6.88	Kimi K2.6 —	Kimi	中国	1000B	开源推理	$0.3653
44.93	-7.09	Claude Opus 4.6 high	Anthropic	美国	large	闭源推理	N/A
44.25	-7.77	Muse Spark —	Meta	美国	—	闭源推理	N/A
43.34	-8.68	GPT-5.2 , medium	OpenAI	美国	large	闭源推理	N/A
43.02	-9.00	Kimi K2.7 Code —	Kimi	中国	1000B	开源推理	$0.2220
42.88	-9.14	MiMo-V2.5-Pro —	Xiaomi	中国	1023B	开源推理	$0.0339
42.29	-9.73	Inkling —	Thinking Machines	美国	975B	开源推理	$0.3389
42.21	-9.81	Hy3 —	Tencent	中国	299B	开源推理	$0.0357
42.12	-9.90	DeepSeek V4 Flash , high	DeepSeek	中国	284B	开源推理	$0.0673
41.87	-10.15	Claude Opus 4.5 —	Anthropic	美国	large	闭源推理	N/A
41.74	-10.28	Nex-N2-Pro —	Nex AGI	中国	397B	开源推理	N/A
41.64	-10.38	Solar Pro 4 —	Upstage	韩国	medium	闭源推理	$0.2231
41.37	-10.65	MiMo-V2-Pro —	Xiaomi	中国	—	闭源推理	N/A
41.22	-10.80	GPT-5.2 Codex —	OpenAI	美国	large	闭源推理	N/A
41.18	-10.84	Inkling Small —	Thinking Machines	美国	266B	开源推理	$0.0735
41.07	-10.95	Qwen3.6 Max Preview —	Alibaba	中国	large	闭源推理	N/A
40.97	-11.05	GLM-5.1 —	Z AI	中国	744B	开源推理	$0.3042
40.94	-11.08	GPT-5.4 mini , medium	OpenAI	美国	large	闭源推理	$0.4948
40.65	-11.37	Grok Build 0.1 0616 —	SpaceXAI	美国	large	闭源推理	$0.2252
40.61	-11.41	Gemini 3 Pro Preview low	Google	美国	large	闭源推理	N/A
40.55	-11.47	GLM-5 —	Z AI	中国	744B	开源推理	N/A
40.49	-11.53	Qwen3.6 Plus —	Alibaba	中国	large	闭源推理	$0.3569
39.85	-12.17	JT-4.1 Flash 236B A21B —	China Mobile	中国	236B	闭源	N/A
39.71	-12.31	GPT-5.4 nano —	OpenAI	美国	medium	闭源推理	$0.1493
39.71	-12.31	Agnes 2.5 Pro Alpha —	Sapiens AI	新加坡	medium	闭源推理	N/A
39.37	-12.65	Qwen3.7 Plus —	Alibaba	中国	medium	闭源推理	$0.2424
39.05	-12.97	GLM-5-Turbo —	Z AI	中国	—	闭源推理	N/A
38.87	-13.15	MiniMax-M2.7 —	MiniMax	中国	230B	开源推理	$0.0777
38.74	-13.28	Gemini 3 Flash —	Google	美国	large	闭源推理	N/A
38.32	-13.70	Nemotron 3 Ultra —	NVIDIA	美国	550B	开源推理	$0.3827
38.04	-13.98	MiMo-V2.5 —	Xiaomi	中国	310B	开源推理	$0.0104
37.95	-14.07	Grok 4.3 Non-reasoning, low, medium	SpaceXAI	美国	large	闭源推理	$0.1454
37.95	-14.07	Grok 4.20 0309 v2 —	SpaceXAI	美国	large	闭源推理	N/A
37.82	-14.20	Ling 3.0 Flash —	InclusionAI	中国	124B	开源推理	$0.0377
37.70	-14.32	Qwen3.6 27B —	Alibaba	中国	27.8B	开源推理	$0.2918
37.47	-14.55	GPT-5.1 —	OpenAI	美国	large	闭源推理	$0.3040
37.44	-14.58	Grok 4.20 0309 —	SpaceXAI	美国	large	闭源推理	N/A
37.44	-14.58	Gemini 3.5 Flash-Lite —	Google	美国	medium	闭源推理	$0.0965
37.43	-14.59	Solar Open2 250B —	Upstage	韩国	250B	开源推理	N/A
37.39	-14.63	Claude 4.5 Sonnet —	Anthropic	美国	large	闭源推理	$0.4643
37.32	-14.70	MiMo-V2-Omni-0327 —	Xiaomi	中国	—	闭源推理	N/A
37.04	-14.98	GPT-5 Codex —	OpenAI	美国	large	闭源推理	N/A
36.02	-16.00	Kimi K2.5 —	Kimi	中国	1000B	开源推理	$0.0964
35.87	-16.15	MiMo-V2-Omni —	Xiaomi	中国	—	闭源推理	N/A
35.60	-16.42	GPT-5.1 Codex —	OpenAI	美国	large	闭源推理	N/A
35.35	-16.67	GLM 5V Turbo —	Z AI	中国	—	闭源推理	N/A
35.31	-16.71	GPT-5 ChatGPT, low, medium, minimal	OpenAI	美国	large	闭源推理	$0.2572
35.06	-16.96	Muse Glimmer —	Meta	美国	30B	开源推理	$0.0732
35.01	-17.01	A.X-K2 —	SK Telecom	韩国	692B	开源推理	N/A

Mid-tier weaker (II 25 – 35)中端弱于（II 25 – 35）

II	Δ	Name名称	Creator厂商	Country国家	Params参数	Type类型	Cost成本
34.60	-17.42	Qwen3.5 27B —	Alibaba	CN	27.8B	openR	N/A
34.54	-17.48	Claude 4.1 Opus —	Anthropic	US	large	closedR	N/A
34.47	-17.55	MiniMax-M2.5 —	MiniMax	CN	230B	openR	N/A
34.46	-17.56	GLM-4.7 —	Z AI	CN	357B	openR	$0.3586
34.60	-17.42	Qwen3.5 27B —	Alibaba	中国	27.8B	开源推理	N/A
34.54	-17.48	Claude 4.1 Opus —	Anthropic	美国	large	闭源推理	N/A
34.47	-17.55	MiniMax-M2.5 —	MiniMax	中国	230B	开源推理	N/A
34.46	-17.56	GLM-4.7 —	Z AI	中国	357B	开源推理	$0.3586

Distribution of all 396 weaker models by II band全部 396 个弱于 Qwen 的模型：按 II 分段分布

II 45-50

II 40-45

II 35-40

II 30-35

II 25-30

II 20-25

II 15-20

II 10-15

II < 10

Final Insights总结洞察

1. Headline for Qwen3.8 27B.1. Qwen3.8 27B 的头条定位。 It is the highest-scoring 27B dense model in the world. Only MoE models with hundreds of billions of total parameters (Claude Opus 5, Claude Fable 5, GPT-5.6 Sol/Terra, Grok 4.6, Kimi K3, GLM-5.3, Qwen3.8 Max, Qwen3.8 2.4T A95B, Claude Opus 4.8, Muse Spark 1.2, GPT-5.5, Gemini 3.7 Flash, Grok 4.5, Claude Sonnet 5, Claude Opus 4.7) reliably beat it. 它是全球得分最高的 27B 稠密模型。只有参数量数百亿的 MoE 模型（Claude Opus 5、Claude Fable 5、GPT-5.6 Sol/Terra、Grok 4.6、Kimi K3、GLM-5.3、Qwen3.8 Max、Qwen3.8 2.4T A95B、Claude Opus 4.8、Muse Spark 1.2、GPT-5.5、Gemini 3.7 Flash、Grok 4.5、Claude Sonnet 5、Claude Opus 4.7）能稳定胜过它。

2. The Chinese open-source ecosystem collided at 52 II.2. 中国开源阵营在 52 II 撞线。 DeepSeek V4 Pro 0813 (53.20), GLM-5.2 (52.64), DeepSeek V4 Flash 0731 (51.77) — all within 2 II of Qwen3.8 27B. The structural difference between top-tier opens and mid-tier closed (Anthropic Sonnet 4.6 at 48.37, GPT-5.2 at 43.34) is now ~10 II, not 20+. DeepSeek V4 Pro 0813（53.20）、GLM-5.2（52.64）、DeepSeek V4 Flash 0731（51.77）——全部在 ± 2 II 之内。顶级开源与中端闭源（Anthropic Sonnet 4.6 在 48.37、GPT-5.2 在 43.34）的差距已缩减到 约 10 II，而非 20+。

3. Anthropic's flagship is heavier than Qwen3.8 27B.3. Anthropic 旗舰跑不过 Qwen3.8 27B。 Claude Opus 4.6 (max, 44.93) is +7 II behind. The premium Anthropic tier (Opus 5 family) lands at 60-63, requiring ~9 II jump to escape Qwen3.8 27B's reach. Claude Opus 4.6（max，44.93）落后 7 II。Anthropic 高端档（Opus 5 家族）落在 60-63，要跳出 Qwen3.8 27B 的射程需要 ~9 II 的跃迁。

4. Open-weight wins above Qwen3.8 27B rating:4. 强过 Qwen3.8 27B 的开源只有 2 个： Only Kimi K3 (59.70) and Qwen3.8 2.4T A95B (57.70). All 14 closed models above Qwen3.8 27B are from Anthropic, OpenAI, SpaceXAI, Google, Z AI, or Alibaba-closed. Kimi K3（59.70）与 Qwen3.8 2.4T A95B（57.70）。其余 14 个在 Qwen 之上的都是 Anthropic、OpenAI、SpaceXAI、Google、Z AI、阿里闭源。

5. Qwen3.8 27B vs Qwen3.x legacy.5. Qwen3.8 27B vs Qwen3.x 历史。 Qwen3.5 27B scores 34.60 — Qwen3.8 27B is +17.42 ahead. Qwen3.6 27B scores 37.70 — Qwen3.8 27B is +14.32 ahead. Qwen3.7 Max (closed) at 46.71 — Qwen3.8 27B is +5.31 ahead. The Qwen3.8 generation absorbed a generational leap. Qwen3.5 27B 得分 34.60 —— Qwen3.8 27B 领先 +17.42。Qwen3.6 27B 得分 37.70 —— Qwen3.8 27B 领先 +14.32。Qwen3.7 Max（闭源）46.71 —— Qwen3.8 27B 领先 +5.31。Qwen3.8 这一代完成了一次代际跃迁。

6. Cost comparison.6. 成本对比。 All comparable top-tier models charge $0.05-$1.50 per task. Qwen3.8 27B has no public per-task cost (only raw token prices: $0.45 in / $3.20 out / $0.05 cache per 1M tokens). For an open-weight model with 47k output tokens / task and reasoning-heavy inference, real-world cost is dominated by the output token rate. 所有可比顶级模型每任务定价 $0.05 – $1.50。Qwen3.8 27B 无公开每任务定价（仅 raw token 价格：$0.45 in / $3.20 out / $0.05 cache 每 1M tokens）。作为单任务 47k 输出 token 的推理型开源模型，实际使用成本主要被输出 token 主导。