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
Intern-S2-Preview
17 atoms · 跨 1 天 · 首见 2026-05-15 · 最近 2026-05-15
三色: 🟦 fact 17 · 🟥 take 0 · stance ▲5/▼0/◆1
来源: X 17
时态: fresh:17
标签: 好信源:14 · 好数字:5
事实 (fact) (17)
- 2026-05-15 · 参数
an efficient 35B scientific multimodal foundation model
tags: 好数字·好信源 · → daily
value: qty=35B
- 2026-05-15 · 科学任务性能
Delivers performance comparable to the trillion-scale Intern-S1-Pro on core scientific tasks
tags: 好数字·好信源 · → daily
value: direction=comparable to Intern-S1-Pro
- 2026-05-15 · 开源特性
The first open-source model with material crystal structure generation capabilities and strong general capabilities
tags: 好信源 · → daily
value: text=first open-source model with material crystal structure generation capabilities
- 2026-05-15 · 科学智能体能力
Significantly stronger scientific agent capabilities on multiple benchmarks
tags: 好信源 · → daily
value: direction=stronger than previous on multiple benchmarks
- 2026-05-15 · 推理效率优化
Improves MTP acceptance rate and token generation speed via shared-weight MTP + KL loss
tags: 好信源 · → daily
value: text=improves MTP acceptance rate and token generation speed via shared-weight MTP + KL loss
- 2026-05-15 · CoT压缩效果
CoT compression shortens responses while preserving strong reasoning, improving both performance and efficiency
tags: 好信源 · → daily
value: text=shortens responses while preserving strong reasoning, improving both performance and efficiency
- 2026-05-15 · 参数规模
35B parameters
tags: 好数字 · → daily
value: qty=35B
- 2026-05-15 · 训练延续起点
continued from Qwen3.5
tags: 好信源 · → daily
value: qty=Qwen3.5
- 2026-05-15 · RL训练使用了共享权重的MTP与KL损失
RL uses shared-weight MTP with KL loss to reduce train-inference mismatch, raise MTP accept rate, and improve token generation speed
tags: 好信源 · → daily
- 2026-05-15 · RL中加入了CoT压缩
CoT compression is added during RL, preserving reasoning quality while shortening responses
tags: 好信源 · → daily
- 2026-05-15 · 文本推理上下文窗口
128K text reasoning context
tags: 好数字 · → daily
value: qty=128K
- 2026-05-15 · 多模态评估上下文窗口
64K multimodal evaluation context
tags: 好数字 · → daily
value: qty=64K
- 2026-05-15 · 默认启用思考模式
thinking mode on by default
tags: 好信源 · → daily
- 2026-05-15 · 支持工具调用
tool calling support
tags: 好信源 · → daily
- 2026-05-15 · 支持的服务框架
LMDeploy, vLLM, and SGLang serving
tags: 好信源 · → daily
- 2026-05-15 · 通过OpenAI兼容端点支持agent集成
agent integration through OpenAI-compatible endpoints
tags: 好信源 · → daily
- 2026-05-15 · 通过LMDeploy的Anthropic兼容端点集成Claude Code
Claude Code integration through LMDeploy’s Anthropic-compatible endpoint
tags: 好信源 · → daily
时间轴 (近 20)
- 2026-05-15 ·
fact · 参数 · → daily
- 2026-05-15 ·
fact · 科学任务性能 · → daily
- 2026-05-15 ·
fact · 开源特性 · → daily
- 2026-05-15 ·
fact · 科学智能体能力 · → daily
- 2026-05-15 ·
fact · 推理效率优化 · → daily
- 2026-05-15 ·
fact · CoT压缩效果 · → daily
- 2026-05-15 ·
fact · 参数规模 · → daily
- 2026-05-15 ·
fact · 训练延续起点 · → daily
- 2026-05-15 ·
fact · RL训练使用了共享权重的MTP与KL损失 · → daily
- 2026-05-15 ·
fact · RL中加入了CoT压缩 · → daily
- 2026-05-15 ·
fact · 文本推理上下文窗口 · → daily
- 2026-05-15 ·
fact · 多模态评估上下文窗口 · → daily
- 2026-05-15 ·
fact · 默认启用思考模式 · → daily
- 2026-05-15 ·
fact · 支持工具调用 · → daily
- 2026-05-15 ·
fact · 支持的服务框架 · → daily
- 2026-05-15 ·
fact · 通过OpenAI兼容端点支持agent集成 · → daily
- 2026-05-15 ·
fact · 通过LMDeploy的Anthropic兼容端点集成Claude Code · → daily
← 实体目录 · 系统日志