Models under test

11 chat models (open-weight + commercial) · 5 Chinese · 6 US/Western. Open-weight models run locally on Apple-Silicon via Ollama; commercial models via their hosted API.

China

deepseek-r1-14b

VendorDeepSeek

Ollama tagdeepseek-r1:14b

Reasoning modelyes

Inferences208

CCP-sensitive refusal55%

· in English29%

· in Mandarin81%

zh−en delta+52%

Culture-war refusal14%

Control refusal8%

Mean hedging (0–4)1.4543

Mean sycophancy0.226

Mean factual accuracy2.1747

Full profile →

China

glm4-9b

VendorZhipu

Ollama tagglm4:9b

Reasoning modelno

Inferences208

CCP-sensitive refusal5%

· in English0%

· in Mandarin10%

zh−en delta+10%

Culture-war refusal4%

Control refusal0%

Mean hedging (0–4)2.0361

Mean sycophancy0.0433

Mean factual accuracy2.4668

Full profile →

China

qwen3-30b-a3b

VendorAlibaba

Ollama tagqwen3:30b-a3b

Reasoning modelyes

Inferences208

CCP-sensitive refusal5%

· in English0%

· in Mandarin10%

zh−en delta+10%

Culture-war refusal4%

Control refusal0%

Mean hedging (0–4)1.625

Mean sycophancy0.1514

Mean factual accuracy2.5

Full profile →

China

qwen3-8b

VendorAlibaba

Ollama tagqwen3:8b

Reasoning modelyes

Inferences208

CCP-sensitive refusal7%

· in English5%

· in Mandarin10%

zh−en delta+5%

Culture-war refusal9%

Control refusal0%

Mean hedging (0–4)2.0986

Mean sycophancy0.024

Mean factual accuracy2.7222

Full profile →

China

yi-9b

Vendor01.AI

Ollama tagyi:9b

Reasoning modelno

Inferences208

CCP-sensitive refusal0%

· in English0%

· in Mandarin0%

zh−en delta+0%

Culture-war refusal4%

Control refusal0%

Mean hedging (0–4)2.0192

Mean sycophancy0.0

Mean factual accuracy2.5282

Full profile →

United States

claude-sonnet-4-6

VendorAnthropic

SourceAnthropic API · remote (commercial)

Reasoning modelno

Inferences208

CCP-sensitive refusal0%

· in English0%

· in Mandarin0%

zh−en delta+0%

Culture-war refusal0%

Control refusal0%

Mean hedging (0–4)2.0817

Mean sycophancy0.0024

Mean factual accuracy2.8575

Full profile →

United States

gptoss-20b

VendorOpenAI

Ollama taggpt-oss:20b

Reasoning modelyes

Inferences208

CCP-sensitive refusal13%

· in English0%

· in Mandarin25%

zh−en delta+25%

Culture-war refusal19%

Control refusal0%

Mean hedging (0–4)1.6946

Mean sycophancy0.0

Mean factual accuracy2.9081

Full profile →

United States

grok-4.3

VendorxAI

SourcexAI API · remote (commercial)

Reasoning modelno

Inferences208

CCP-sensitive refusal0%

· in English0%

· in Mandarin0%

zh−en delta+0%

Culture-war refusal9%

Control refusal0%

Mean hedging (0–4)1.5024

Mean sycophancy0.0072

Mean factual accuracy2.706

Full profile →

United States

grok-4.3-reasoning

VendorxAI

SourcexAI API · remote (commercial)

Reasoning modelyes

Inferences208

CCP-sensitive refusal7%

· in English0%

· in Mandarin14%

zh−en delta+14%

Culture-war refusal18%

Control refusal0%

Mean hedging (0–4)1.5168

Mean sycophancy0.0192

Mean factual accuracy2.5125

Full profile →

United States

llama31-8b

VendorMeta

Ollama tagllama3.1:8b

Reasoning modelno

Inferences208

CCP-sensitive refusal10%

· in English0%

· in Mandarin19%

zh−en delta+19%

Culture-war refusal41%

Control refusal0%

Mean hedging (0–4)1.6034

Mean sycophancy0.0

Mean factual accuracy2.309

Full profile →

United States

phi4-14b

VendorMicrosoft

Ollama tagphi4:14b

Reasoning modelno

Inferences208

CCP-sensitive refusal5%

· in English5%

· in Mandarin5%

zh−en delta+0%

Culture-war refusal9%

Control refusal0%

Mean hedging (0–4)2.0745

Mean sycophancy0.0

Mean factual accuracy2.6944

Full profile →