Kimi K3 Launches with 2.8 T Parameters – A New Milestone for Chinese Open‑Source Models

Last week rumors suggested Kimi might reach a trillion parameters; the K3 model actually ships with 2.8 T parameters, native visual understanding and a 1 M token context window, pushing Chinese open‑source models to a globally competitive scale.

Is this the last top‑tier open‑source model?

K3’s total parameters are 2.8 T. Over the past year the Kimi series set the open‑source size ceiling (K2 at 1 T, K2.5 as the first multimodal model). K3 now ranks among the few worldwide models above 1 T that remain competitive.

Horizontal comparison : K3 should be compared with Fable 5, Opus 4.8, Sol 4.5, GPT‑5.5, GLM 5.2 and others.

K3 achieves state‑of‑the‑art results on programming benchmarks (DeepSWE, FrontierSWE, Terminal Bench 2.1, Program Bench, SWE Marathon) and on agent benchmarks (GDPval‑AA v2, Automation Bench, JobBench, SpreadsheetBench 2, BrowseComp). Overall it outperforms Claude Opus 4.8 and GPT‑5.5, while still lagging behind Claude Fable 5 and GPT‑5.6 Sol; on some engineering tasks it approaches Fable 5 and surpasses GPT‑5.6 Sol.

Pricing remains affordable: with cache hit the cost is ¥2 per million tokens, without cache ¥20, and output cost ¥100. Claude Opus 4.8 costs up to $25 for the same usage.

Why “bigger” is not trivial

The difficulty is not merely stacking parameters but keeping the model runnable and stable. K3 introduces a new low‑level architecture called Kimi Delta Attention (KDA) together with Attention Residuals (AttnRes) to improve information flow across ultra‑long sequences and deeper layers.

K3 also adopts a more aggressive Mixture‑of‑Experts sparsity: 896 experts with only 16 active at a time, using a framework named Stable LatentMoE. This design aims for stable reasoning on long‑context and multi‑turn dialogues. Compared with K2, the combined architectural and training changes yield roughly 2.5× scaling efficiency, meaning the same compute delivers more intelligence.

Impressive effects

The innovations translate into noticeable performance gains.

Using K3 to improve its own codebase, the model can analyze bottlenecks, propose optimizations, modify code, run comparative experiments, and continue iterating after failures—moving beyond simple code‑generation to participating in genuine AI research workflows.

In scientific research scenarios K3 reproduces results such as the I‑Love‑Q universal relation for dense stars within a general‑relativity framework, GWTC‑5 gravitational‑wave analysis, and pulsar timing population studies, showing it can handle rigorous mathematical derivations and data processing.

It also powers 3D/game demos that approach AAA‑quality visual effects.

Real‑world feel

In the author’s first‑hand test, a recent “self‑evolving agent” example from Andrej Karpathy was run on K3. With abundant compute and no contention, the case demonstrated the model’s comprehensive capabilities.

/goal "构建一个名为 Skill Miner 的自动化技能挖掘系统，用于持续发现、提炼、并沉淀'输入一个链接，自动生成一支完整短视频'相关的可复用AI工作流，最终验证这套技能是否真的能产出可用的视频。"

═══════════════════════════════════
【第一部分：前置配置】
═══════════════════════════════════

目标领域：只关注'输入一个URL/链接，自动生成一支完整短视频'相关的开源项目与工作流，
具体包括但不限于：网页/文章内容抓取与提炼、脚本/分镜生成、TTS配音、字幕动画、
素材/转场自动匹配、背景音乐适配、竖屏短视频渲染与合成的开源仓库

挖掘目的：沉淀一套'一个链接 → 一支利于传播的短视频'的完整技能包，
补充到我自己的 skills 目录，之后拿真实文章/链接直接调用生成短视频

验证方式：技能包挖出来后，实际用它处理以下3个不同类型的真实链接，生成3支视频，
并对每支视频做独立质量评分，判断是否比手动/零散调用方式更省事、产出是否可直接使用：
1. https://mp.weixin.qq.com/s/yEWfTLs9W3-QblvPbrGUPQ
2. https://mp.weixin.qq.com/s/TI1PhbO1xm455-WEOBGi2Q
3. https://mp.weixin.qq.com/s/Nra1BerW5kSDZEipJcS6ow

你自己维护的技能库仓库路径：[待补]

═══════════════════════════════════
【第二部分：技能挖掘流水线，共8个独立阶段】
═══════════════════════════════════

每个阶段只做一件事，输出结构化JSON传给下一阶段。

1. Scout：按【目标领域】监控 GitHub Trending 和关键词搜索
   （如 'text-to-video' 'article to video' 'auto video generation' 'content to reels' 等）
   - 搜索条件：stars > 100，最近3个月有更新
   - 只输出候选仓库列表（name/stars/language/last_updated），不做判断

2. Filter：用确定性规则（非LLM）过滤掉明显不符合'链接生成视频'领域的仓库
   （如纯视频剪辑工具但不涉及内容提炼、纯TTS库但不涉及完整流程）
   - 输出 KEEP/REJECT + 理由

3. Reader：按 README → docs/ → examples/ → 源码 的顺序增量读取
   - 前面信息足够时不再往下走，记录 context_loaded 和是否读了源码

4. Workflow Extractor：从上下文中提炼'链接→视频'相关的可复用工作流，标准化为
   {skill_name, goal, inputs, steps, outputs, failure_modes} 结构，
   并标注这个技能对应链路里的哪一步（内容提炼/脚本生成/配音/字幕/转场/合成）
   - 不符合领域或提炼不出来的直接淘汰

5. Skill Score：用固定规则（非LLM）打分，检查项至少包括：
   有README依据、有示例、步骤数达标、可复用、非仓库专属逻辑、
   确实能拼进'链接→视频'完整链路；任一必选项不过直接拒绝

6. Skill Generator：转成标准化技能包，目录结构：
   skill-name/SKILL.md, examples.md, commands.md, metadata.json, tests.md

7. Reviewer：独立于Generator的审核agent，只回答
   '这个技能能否不做修改直接拼进链接生成视频的流程里使用'，YES/NO + 一段理由

8. Publisher：审核通过后，在【技能库仓库路径】创建分支、提交技能包、准备PR说明
   （不自动合并，等待人工审批；不允许对第三方仓库发起任何写操作或PR）

【流水线验收标准，跑完逐条核对】
1. Scout 发现候选仓库 ≥30 个
2. Filter 明确记录 KEEP/REJECT 数量及理由分布
3. Reader 至少 60% 的仓库无需读源码即可判断（记录具体比例）
4. Workflow Extractor 提炼出候选工作流 ≥5 个，且全部标注对应链路环节
5. Skill Score 每个工作流给出完整 checks 明细，不能只给最终分数
6. Skill Generator 完整生成 ≥2 个技能包（5个文件齐全）
7. Reviewer 审核与 Generator 独立执行，每个技能包有独立 YES/NO + 理由
8. Publisher 至少 1 个技能包通过审核，生成完整 PR 草稿并创建本地分支

【执行注意事项】
- 分阶段输出过程进度，每阶段结束后简要说明做了什么、筛掉/通过了多少
- 中途遇到GitHub API限流、仓库读取失败、视频渲染失败等问题，自主定位并调整策略继续推进，不要卡住
- 使用合适的SKILL完成相关任务，避免重复造轮子
- Publisher阶段严禁对Scout/Reader读取的任何第三方仓库发起写操作或PR

【输出格式】
分阶段输出过程进度
最终输出：完整代码 + 技能包文件 + SKILL_PIPELINE.md + 3支生成的视频 + report.html + PR草稿内容" --max-iterations 100 --completion-promise "DONE"

The pipeline successfully completed all eight stages, discovered over 150 video‑generation projects, and produced a full report, confirming that K3 can drive complex, multi‑stage autonomous workflows.

Summary

One‑sentence takeaway: Sonnet‑level pricing with Opus‑level experience.

K3 may not universally surpass Opus 4.8, but in everyday programming workflows it offers a cost‑effective, reliable alternative that matters more to developers than raw benchmark scores.

The real test of its longevity lies in the underlying linear‑attention architecture and day‑to‑day usability; combined with KimiCode, K3 is currently the best domestic substitute for Claude + ClaudeCode.

Join the discussion by replying “join” to the public account.