以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

MAI-Thinking-1

15 atoms · 跨 2 天 · 首见 2026-06-02 · 最近 2026-06-09

三色: 🟦 fact 10 · 🟥 take 5 · stance ▲1/▼0/◆7

来源: X 15

时态: fresh:15

标签: 好数字:9 · 好观点:4 · 好信源:4

2026-06-02 · GPQA 分数
MAI-Thinking-1 has weirdly low GPQA & Terminal Bench 2.0 scores
tags: 好数字 · → daily
value: qty=weirdly low · date=H1 2026
2026-06-02 · Terminal Bench 2.0 分数
MAI-Thinking-1 has weirdly low GPQA & Terminal Bench 2.0 scores
tags: 好数字 · → daily
value: qty=weirdly low · date=H1 2026
2026-06-02 · 109 页技术报告可用
the 109 page tech report is gold
tags: 好信源 · → daily
2026-06-09 · primary motivation was hedge and leverage against Ant
I'm utterly convinced that primary motivation for MAI-Thinking-1 was as a hedge and leverage vis-à-vis Ant pulling these kind of stunts.
tags: 好观点 · → daily
2026-06-02 · 性能低于 Meta Spark
Stats below Meta Spark, though
tags: 好观点 · → daily
value: direction=below
2026-06-02 · 模型性能接近 Sonnet 级别
MAI-Thinking-1 by Microsoft looks to be approaching sonnet level model
tags: 好观点 · → daily
2026-06-02 · 预训练策略与 phi 系列相反
which is exact opposite of what they were doing with phi models
tags: 好观点 · → daily

2026-06-02 · HealthBench Professional 基准表现
they've reported health evals: HealthBench Professional
tags: 好数字·好信源 · → daily
2026-06-02 · MedXpertQA 基准表现
they've reported health evals: MedXpertQA
tags: 好数字·好信源 · → daily
2026-06-02 · active parameters
It’s a 35B active parameter MoE with a 256K context window.
tags: 好数字 · → daily
value: qty=35B
2026-06-02 · context window
It’s a 35B active parameter MoE with a 256K context window.
tags: 好数字 · → daily
value: qty=256K
2026-06-02 · AIME 2025 score
it’s achieved 97% on AIME 2025, the key measure of its general-purpose reasoning abilities.
tags: 好数字 · → daily
value: qty=97%
2026-06-02 · SWE Bench Pro score
It's at 53% on SWE Bench Pro, placing it right alongside Opus 4.6 on one of the toughest coding benchmarks.
tags: 好数字 · → daily
value: qty=53%
2026-06-02 · 附带技术报告
they have a detailed tech report too!
tags: 好信源 · → daily
value: date=2026-06-02
2026-06-02 · 预训练使用 29T 唯一 token 且无合成 token
they got 29T unique tokens without any synthetic tokens for pretraining
tags: 好数字 · → daily
value: qty=29T tokens

2026-06-09 · narrative · primary motivation was hedge and leverage against Ant · → daily
2026-06-02 · fact · active parameters · → daily
2026-06-02 · fact · context window · → daily
2026-06-02 · fact · AIME 2025 score · → daily
2026-06-02 · fact · SWE Bench Pro score · → daily
2026-06-02 · narrative · GPQA 分数 · → daily
2026-06-02 · narrative · Terminal Bench 2.0 分数 · → daily
2026-06-02 · narrative · 性能低于 Meta Spark · → daily
2026-06-02 · fact · 附带技术报告 · → daily
2026-06-02 · fact · HealthBench Professional 基准表现 · → daily
2026-06-02 · fact · MedXpertQA 基准表现 · → daily
2026-06-02 · narrative · 模型性能接近 Sonnet 级别 · → daily
2026-06-02 · narrative · 109 页技术报告可用 · → daily
2026-06-02 · fact · 预训练使用 29T 唯一 token 且无合成 token · → daily
2026-06-02 · narrative · 预训练策略与 phi 系列相反 · → daily