可能是最好用的免费模型了——Agnes 2.5 flashProbably the Best Free Model — Agnes 2.5 flash


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/UygkZyDqWnBVl0Ts0Q0lAw
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

2026 年 7 月 13 日,新加坡 Sapiens AI 正式发布了 Agnes-2.5-Flash——一个定位为"全能力文本模型"的机型。同日上线的还有 Agnes Code 桌面端,以及尚未正式发布但已在内部完成训练的旗舰版本 Agnes-2.5-Pro。

看完刚才这段话,可能你还是一头雾水,什么AI?什么Flash?没听说过。

不过“免费”二字,倒是很有含金量。毕竟免费意味着无限试错、无限迭代,既不用绞尽脑汁写一些文言文prompt节省token,也不用在月底为账单感到心痛。免费模型无论是给老玩家补充token弹药,还是帮新手小白探索Agent玩法,都是不错的挑选。

尽管如此,如果模型的能力太差,听不懂需要、多次返工,就算免费也会令人抓狂。幸运的是,看能力,Agnes已经接近了全球第一梯队。


一、Agnes 模型概况

1.1 什么 AI ?

Agnes AI 是新加坡 AI 公司 Sapiens Technology Pte. Ltd. 旗下的机型品牌。这家公司成立于 2024 年(一说 2025 年 7 月),创始人与 CEO 是 Bruce Yang——新加坡莱佛士书院校友,NUS 人工智能方向博士生。团队核心成员来自 NUS、NTU、MIT、斯坦福、UC Berkeley、UT Austin 等高校。

Sapiens AI 的融资情况:

轮次
时间
金额
领投方
Seed
2025 年
未公开
—
Series A
2026 年 2 月 13 日
1000 万美元
LOOK FORWARD VCC
累计融资
—
约 2000 万美元
—

公司定位格外明确:自己训练模型,不用第三方 API。它不是 OpenRouter 那种模型聚合平台,而是从底层架构到应用层全部自研。这也是为什么它能做到"不限期免费"——成本结构完全掌握在自己手里。

截至 2026 年 6 月,Agnes AI 已有超过 600 万使用者,每周处理超过 5.41 万亿 tokens(其中文本 token 3.25 万亿、多模态 2.16 万亿)。这个数字意味着什么?大概相当于每周免费输出数百万本长篇小说的文字量。

1.2 什么 Flash ?

Agnes AI 的机型矩阵围绕三类模型展开:

文本 / Agent 模型

模型
发布时间
定位
状态
agnes-1.5-flash
早期
快速对话、内容生成、简单多模态任务
✅ 已发布
agnes-2.0-flash
2026 年 5 月前
编码、推理、Tool Calling、Agent 工作流、图像理解
✅ 当前主力
agnes-2.5-flash2026 年 7 月 13 日
新一代编码与 Agent 主力模型,全面升级
✅ 已发布
agnes-2.5-pro
预计 2026 年 7 月下旬
旗舰通用推理,对标 Claude Opus 4.8 / GLM-5.2
🚧 Pro 预览版

图像模型

模型
用途
agnes-image-2.0-flash
文生图、图生图、快速出图
agnes-image-2.1-flash
高精细图生图、图像编辑

视频模型

模型
用途
agnes-video-v2.0
文生视频、图生视频、关键帧动画

另外还有创意平台 Pavo(web 端 AI 创作工作室)和桌面 IDE 工具 Agnes Code,统一账号体系,一套积分通吃所有能力。

1.3 Agnes-2.5-Flash 的核心规格

依据官方文档和发布信息,Agnes-2.5-Flash 的关键参数如下:

维度
数值
上下文窗口
256K tokens
最大输出
65K tokens
多模态输入
文本 + 图像 URL
Thinking Mode
兼容(编码/调试情形启动深度思考)
Tool Calling
兼容(Function Calling)
Stream 输出
兼容
API 兼容
OpenAI SDK + Anthropic thinking 字段兼容
定价
输入 0(永久免费)

比较前一代 agnes-2.0-flash,官方强调在以下几个维度有显著提高:

1.4 与 OpenAI / Anthropic 的兼容性

显然兼容。Agnes AI 给出的是 OpenAI 兼容 API:

from openai import OpenAIclient = OpenAI(    base_url="https://apihub.agnes-ai.com/v1",    api_key="your-key-here")response = client.chat.completions.create(    model="agnes-2.5-flash",    messages=[{"role": "user", "content": "你好"}],    stream=True)

同样兼容 Anthropic 风格的 thinking 字段。这意味着如果你的项目已经在用 OpenAI 或 Anthropic SDK,只需要替换 base_url 和 model 名称就能切换,集成成本极低。

1.5 免费?

很多人会问:免费能做到什么程度?是不是有隐藏限制?

实际情况是:

项目
说明
费用
输入、输出均 $0
有效期
不限期(官方承诺)
RPM 限制
免费版 30 RPM(实际可执行约 20 RPM)
模型挑选
只能使用 Agnes AI 自有模型
第三方模型
需要借助 Agnes Code 等平台,按会员收费(定价未公布)

也就是说:Agnes AI 自己的模型全部免费,涵盖文本、图像和视频。如果你借助 OpenRouter 等第三方平台间接使用,那是另一回事。RPM 对日常开发够用,但如果你要跑大规模自动化 Agent,可能需要升级到 Token Plan 或企业计划。


二、工艺创新点

上一章笔者了解了 Agnes 的"硬指标":免费、免费 API、免费多模态。但真正让这套打法跑通的,不是营销话术,而是一整套从底层训练到上层调度的工艺栈。

下面从最简单的直觉出发,一步步拆解 Agnes 有哪些自己独有、别人没有的工艺优化。


2.1 强化学习的"防崩溃"算法:DSPO

先回答一个根本难题:模型怎么变聪明?

大模型训练分几步:先用海量文本做预训练(读万卷书),再用人类反馈微调(学习规范),最后一步是强化学习(RL)——让模型在复杂任务中借助"试错+奖惩"学会最优策略。

但 RL 有个致命弱点:复杂任务中的奖励信号太稀疏、太高频的失败会导致策略崩溃。

举个例子。让 AI 帮你做一份研究报告,流程大概是:理解需要 → 搜索资料 → 整理分析 → 写结论。如果最后只给一个"做完了就给奖励"的信号,模型根本不知道哪一步做得好、哪一步做错了。更糟的是,在搜索情形中,大量尝试会失败,奖励信号几乎为零,梯度变得不稳定,训练直接崩塌。

这就是 DeepSeek 的 GRPO 等主流算法碰到的难题——它们在简单推理任务上表现出色,但在需要多轮搜索、工具调用、长链规划的 Agent 任务中容易"学崩"。

Agnes 的做法叫 DSPO(Dynamic-filter Sequence-level Policy Optimization),中文名"动态过滤序列级策略优化"。它不是又发明了一个新 RL 算法,而是把两个已知的好思想巧妙组合在一起。

先看第一个难题:优化单位不匹配。 GRPO 在 token 级别做优化,但奖励是对整个序列给的——相当于你用每个字的表现来打分,但评判标准却是整句话写得好不好。这就像用每一步来评价一场马拉松的成绩。DSPO 改成序列级优化:不再盯着每个 token 单独算梯度,而是对整条轨迹算一个综合校正因子。这样梯度方差大幅降低,训练不再崩溃。

第二个难题:稀疏奖励浪费样本。 搜索 Agent 的训练过程中,很多 prompt 的 trial group 要么全部成功、要么全部失败,这些组的长处信号为零或不存在,白白浪费了算力。DSPO 引入了动态过滤机制:只保留"既有成功又有失败"的 trial group 用于训练,强制确保每个 batch 都有非零的、有效的长处信号。

一个比喻就好理解了:

这不是一个小改进。依据 Sapiens AI 在 arXiv 上发表的论文 DSPO,他们在 7B 模型上训练出的 Agent,比之前的基线模型(同规模 GRPO)提高了 34.1%,还超越了一个14B 模型在近 9% 的相对提高——一个参数量只有对方一半的模型,击败了更大的对手。

对使用者的感知价值是什么?模型更稳了,Agent 任务的可靠性更高了,搜索和推理不再是"碰运气"而是"稳定输出"。

2.2 CodeAgents:用"伪代码"而不是自然语言来做 Agent 协作

前面说了如何让单个模型变强,现在聊一个不同的难题:多个 AI Agent 之间怎么高质沟通? 当下使用多Agent拆分复杂任务,分工合作已经有各种细致的落地应对思路,不过Agnes在此基础上,开展了新的优化。

假设你有一个研究 Agent 平台,里面有三个角色:

如果用自然语言让这三个 Agent 协作,会发生什么? Planner 说"帮我查一下东南亚科技投资趋势",Solver 搜了一堆信息,然后 Replanner 看了结果觉得不对,又要 Planner 重新调整方向。难题是:每个人的语言理解是模糊的。"查一下"可以理解为深度报告,也可以理解为三句话总结。"科技投资"可以指风投、并购、政府基金……歧义越多,返工越多。

CodeAgents 的核心思路是:不要用自然语言做 Agent 之间的通信协议,改用结构化伪代码。

具体怎么做?

把 Planner、Solver、Replanner 的交互编码成类似这样的结构:

task:"analyze SE Asia tech investment trends"plan:-step:"search"tool:"google_search"query:"Southeast Asia tech funding 2026"expected_length:">= 5 results"-step:"filter"condition:"result_count >= 3"subroutine:"summarize_trends"

这不是真的 Python 代码,而是一种类型化的伪代码格式:

这就像软件工程中 API 的约定:前端和后端借助 JSON schema 通信,而不是互相发微信语音。

这种格式化的处理思路有一点像 MCP。 MCP 是 AI 应用连接外部工具和数据源的协议标准(例如怎么调用搜索引擎、数据库、API),而 CodeAgents 是 Agent 平台内部的编排框架——它管的是 Planner、Solver、Replanner 这些角色之间怎么交换信息,不管怎么接入外部工具。

换句话说,MCP 应对的是"AI 怎么用工具",CodeAgents 应对的是"多个 AI 之间怎么高质协作"。在实际机型中,两者可以共存:内部多 Agent 通信走 CodeAgents 式的结构化编排,外部工具调用可以借助 MCP 来对接。

这个外观带来了什么?

指标
比较自然语言 baseline
输入 token 减少
55–87%
输出 token 减少
41–70%
规划准确率提高
绝对值 +3 到 +36 个百分点
VirtualHome 成功率
56%(新 SOTA)

数据来源:CodeAgents 论文的 arXiv - 基准测试 链接:2507.03254。

注意这里和之前"40-70% token 节省"的粗略说法不同——论文原文的数据范围更大,真实数字是 **输入端节省 55-87%,输出端节省 41-70%**。

当然,这也意味着同样的模型能力下,Agnes 可以服务更多使用者而不增加同等比例的推理成本。这就是"免费"背后的真实数学:不靠补贴,靠省下来的 token。

2.3 Universal Verifier:给"主观题"打分

Bruce Yang 本人将其称为 "secret sauce",未开源。

Benchmark 分数高≠实际感受好。你可以在数学题上拿第一,但在写 PPT、做研究报告、甚至"在群聊里该不该说话"这种主观任务上一塌糊涂。Agnes 的核心机型情形恰恰是这类没有标准答案的任务。

例如:

传统做法是给这些主观任务找人工标注员打分,效率低且不可扩展。

Agnes 的做法是 Universal Verifier(UV):一个内部评估平台,能对主观任务给出评分和奖励信号。

按 Bruce Yang 本人的说法:"否则模型不知道往哪个方向改进。它甚至不知道该往哪儿走。"

UV 的工作方式大致是两层:

第一层:训练过程中的评估器。 对主观型任务,UV 给出生成结果的评分和反馈,告诉模型"哪里做得好、哪里需要改"。它不仅看最终答案对不对,还要看推理路径是否合理。

第二层:上线后的持续监控。 UV 在后台跟踪真实情形中的反馈信号,如果发现某类任务的回答率下降,就触发重新训练。

传统的科研→发布→上线→发现难题的循环太慢。UV 把它变成了一个持续的闭环:

研究 → 评估 → 训练 → 上线 → 反馈  ↑                        ↓  ←←←←←←←← 持续循环 ←←←←←←←

Yang 的原话是:"UV 是笔者的 secret sauce——笔者没有开源它。"它没有公开细节,但从其能力定位来看,它是连接 benchmark 和真实使用者感受的关键桥梁。

2.4 生成平台的"中间层":Intent 翻译而非 Agent 创新

坦率地说,**"三个 Agent 协作"本身不是新概念。** 从 GAN 的 Generator vs Discriminator,到 RLHF 的 Response vs Reward Model,再到各种框架里的 Solver/Reviewer/Replanner 分工,业界早就这么做了。

这里要写清楚的不是"谁用了三个 Agent",而是一个更具体的难题:

当使用者输入是模糊的自然语言,但下游生成模型(图像/视频 Diffusion)需要精确的结构化指令时,中间该放什么?

Agnes 的做法是加一个 Intent Understanding Agent,放在使用者和 Generation/Judge Agent 之间。核心逻辑很简单:

所以这个"三 Agent"的真正价值不在于 agent 数量,而在于LLM 作为意图翻译层,把自然语言的模糊性转化为生成模型可执行的确定性指令。它应对的不是"如何组织多个 Agent"这个新难题,而是"如何让文本生成模型和图像/视频生成模型在同一个工作流中衔接"这个工程难题。

用一句话总结:这不是一个新的创新点,而是一个针对特定情形(多模态生成)的合理工程架构。 值得记录的只是它被明确命名和机型化了。

多Agent开发并不是新鲜事物,不过如何能标准化的使用这种思想,而不依赖于使用者自己人肉Harness,甚至不需要使用者做任何事情,并且要保证稳定性与成效。可以优化到这种程度,那才算是真正值得注意的创新,才不算是似是而非的自我包装。

2.5 自研模型路由:不是所有难题都值得用最大模型

Agnes 不只是一个模型,而是一套模型家族。从 7B 到 30B+ 的参数规模,针对不同任务有不同的小模型专化版本。

Yang 的表述很清楚:

"笔者不需要在每个维度都超过 GPT-4,笔者只需要在特定类别上超越它——研究、PPT 生成、群聊辅助。笔者的平台中每种模式都有专门的模型。"

路由逻辑如下:

使用者请求
可能路由到
"明天天气如何"
小型快速模型
"生成 PPT"
PPT 专化模型
"图像编辑"
Image 2.0/2.1 Flash
"文生视频"
Video V2.0
"复杂代码调试 + Agent 工作流"
Agnes-2.5-Flash / Pro

这本质上是多模型任务路由,类似于 MoE 中"路由器"的思路,但 Agnes 的路由粒度更粗——它不是在模型内部挑选专家,而是在模型之间挑选调度对象。

对成本的直接影响:约一半的使用者流量被路由到小模型或专化模型上,只有一小部分请求需要消耗最大计算资源。这意味着 Agnes 的成本结构中,大头不是"每次调用都用旗舰模型",而是按需分配。这也是"免费"的经济基础之一。

2.6 真实使用者数据飞轮

最后一环:数据的来源。

Agnes AI 有 600 万使用者,日活 20 万以上。这意味着每天有海量真实的、多样的、甚至包含非主流语言和小语种的使用数据被产生。

训练流程大致是:

这就是一个正循环:

使用者越来越多 → 数据越来越多元 → 模型越来越贴使用者需要 → 感受越来越好 → 使用者越来越多

这个飞轮为什么 Agnes 能做到? 因为聚合平台(例如 OpenRouter)只转发请求,看不到使用者和模型之间的完整交互闭环。Agnes 自己训练模型、自己给出 API、自己积累数据,这个闭环是完整的。

2.7 Speed as Product:一分钟内必须完成

贯穿以上所有工艺的最终原则:速度就是机型本身。

Yang 给出的硬标准是:使用者可以在一分钟内完成一次完整的研究。 为什么是一分钟?心理学研究表明,使用者对等待的感受是非线性的:

等待时长
感受
3 秒内
"即时"
10 秒内
"有点慢"
30 秒
注意力开始分散
60 秒
大部分耐心耗尽

要达到这个速度,需要前面的每一项工艺协同配合:自研模型降低推理成本、DSPO 提高训练稳定性(减少反复试验)、CodeAgents 节省 token、意图翻译层减少生成返工、模型路由避免"杀鸡用牛刀"、数据飞轮持续提高成效。

用一句话总结:**"Speed as Product"不是机型上线后再去优化的工程目标,而是从一开始就指导整个工艺栈外观的机型哲学。**

2.8 工艺护城河总结表

工艺维度
核心创新
与传统思路比较
论文/证据来源
自研全栈模型
Text/Vision/Video 全自研,不做聚合
聚合 vs 自产
Coruzant
DSPO
序列级优化 + 动态过滤防崩溃
token 级 GRPO 在 Agent 任务上不稳定
arXiv:2510.09255
CodeAgents
伪代码结构化多 Agent 通信
自然语言 Agent 间通信冗余、易歧义
arXiv:2507.03254
Universal Verifier
主观任务内部评分与持续评估
仅依赖 benchmark 无法衡量真实质量
Coruzant
意图翻译层
LLM 将模糊使用者需要翻译为生成模型的精确指令
多 Agent 架构并非新概念,价值在于应对自然语言→Diffusion 的桥接难题
TechLead Journal #246
多模型路由
7B–30B+ 按任务分发,半流量走小模型
固定用一个旗舰模型处理所有请求
AI Magazine
RLAF + 数据飞轮
真实使用者数据 → 预训练 → 蒸馏 → DSPO
静态数据集训练
Sapiens AI Case Study
Speed as Product
一分钟内完成研究
先保质量再加速
Coruzant

三、表现横评:免费模型能跟谁打?

光说"免费"没意义,人们最关心的还是:它到底有多强?

3.1 Claw-Eval:Agent 能力的试金石

Claw-Eval 是当下最具参考价值的 Agent 基准之一。它评估的是 AI 能否像真正的软件工程师一样,在真实环境中自主规划、调用工具、执行多步骤任务。

排名
模型
Pass@3
1
Claude Opus 4.6
70.4%
2
Step 3.7 Flash
68.3%
3
Claude Sonnet 4.6
67.8%
...
...
...
9Agnes-2.0-Flash60.9%

Agnes-2.0-Flash 以 60.9% Pass@3 排名全球第 9,超越 Gemini 和 MiniMax,进入了全球前十。而 Agnes-2.5-Flash 在这个基准上的具体分数官方尚未公开,但依据官方说法,2.5 版本在所有基准上全面超越了 2.0。

这个排名的含金量很高:Claw-Eval 用的是三次独立测试的 Pass@3 指标——不是跑一次就给你个平均分,而是要三次都能成功才算借助。这意味着模型不仅要"偶尔能做对",还要稳定可靠。

3.2 SWE-bench Verified:软件工程真实任务

SWE-bench Verified 衡量的是模型修复真实 GitHub issues 的能力,被业界视为"代码能力黄金标准"。

官方披露的数据:

模型
SWE-bench Verified
说明
Agnes 2.5 Pro
~82.7
内部评估,已接近或达到顶尖水平
GPT-5.5
~84.4 / 88.7*
闭源旗舰
DeepSeek V4 Pro
80.6
开源旗舰
Hy3
78.0
开源 MoE

*GPT-5.5 在不同评估体系下有两个版本:标准版约 84.4,Codex CLI 版约 88.7。

Agnes 2.5 Pro 的 82.7 分已经格外接近 GPT-5.5。但注意,这是 Pro 版本的成绩——Flash 版本的绝对排名官方没有给出具体的公开数据。依据 LinkedIn 和多家媒体的报道,Flash 在多项基准上超越了 GLM 5.2 和 DeepSeek 54 Pro 的部分子项,总体与同级别的开源模型处于同一竞争区间。

3.3 Multilingual SWE:多语言编码

模型
Multilingual SWE
Agnes 2.5 Pro
~78.7
其他模型
官方仅展示了相对排名,无具体分数

3.4 与主要竞品的横向比较

维度
Agnes-2.5-Flash
DeepSeek V4 Flash
GLM-5.2
Kimi K3
GPT-5.5
Claude Opus 4.8
售价免费
0.07
0.28
0.64
30
75
上下文
256K
128K
128K
1,000,000
1,000,000
1,000,000
开源
❌
✅ Apache 2.0
✅ Apache 2.0
✅ 计划中
❌
❌
Agent 能力
Top 10
强
强
极强
强
极强
编程能力
第一梯队
第一梯队
第一梯队
第一梯队
第一梯队
第一梯队
多模态
文本+图像输入
文本为主
文本+视觉
文本+图像+视频
文本+图像+视频
文本+图像+视频
免费可用
✅ 永久免费
❌ 付费 API
❌ 付费 API
❌ 付费 API
❌ 付费 API
❌ 付费 API

几个观察:

3.5 性价比计算

假设你每天写代码 4 小时,平均每小时产生约 10,000 tokens 输出:

项目
每天消耗
月消耗
GPT-5.5 售价
Claude 售价
Agnes 售价
输出 tokens
~240K
~7.2M
$216
$540
$0
输入 tokens
~480K
~14.4M
$72
$216
$0

每个月光这一个数字,Agnes 就能帮你省下不少钱。当然,这不是说你应该完全依赖免费模型,但至少对于日常开发、原型验证、学习探索,它完全可以信赖,作为主力模型使用。


四、总结:值不值得用?

我自己主要在OpenCode的桌面版本,搭配第三方API使用各家模型,当然Agnes也是其中之一。

在使用者意图理解,代码生成、工作流优化、代码难题排查与应对、文档书写与事实核查几个维度的感受,我对于Agnes 2.5 Flash的评价是,高于Deepseek V4 Flash,与Deepseek V4 Pro相当或略低,不如Hy3但差距不大。

考虑到免费与稳定性(我使用Nvidia给出的免费Deepseek V4 Pro,但其额度很小且本身不稳定),再综合考量表现,我自己已经完全将Agnes 2.5 Flash作为默认的模型使用,偶尔用一下Deepseek V4 Flash开展交叉验证。

合适的情形

不合适的情形

一句话

Agnes-2.5-Flash 是当下市面上唯一一个把"免费 + 第一梯队能力 + 完整多模态"三件事同时做到的机型。它可能不是每一项都拿第一,但在"花最少的钱取得足够好的感受"这件事上,当下没有对手。


参考来源

官方文档与发布

学术论文

基准排名

工艺分析与采访

Benchmark 数据新闻


尾声

Agnes AI 的故事很有趣。一家新加坡创业公司,2000 万美元融资,不做聚合平台,自己训练从文本到图像到视频的全栈模型,然后全部免费开放。

这种打法在任何行业都算激进——但它之所以敢这么干,是因为它有别家没有的东西:完整的自研模型栈。不需要向 OpenAI、Anthropic、Google 交"过路费",所以它能把售价打到零。

免费从来不是终点,而是一个让使用者先尝到甜头的入口。但如果 Agnes-2.5-Pro 能像宣传的那样逼近 Claude Opus 4.8,那这个入口可能比很多人想象的要深遂得多。

至于它能不能真正留住使用者?这种模式可以跑多久?我也没有答案,但也许答案就在笔者的手里。

Source: https://mp.weixin.qq.com/s/UygkZyDqWnBVl0Ts0Q0lAw For copyright issues, please contact: 13951431913#139.com or submit a comment on the right.

On July 13, 2026, Singapore-based Sapiens AI officially released Agnes-2.5-Flash — a model positioned as a "fully-capable text model." Also launched on the same day were Agnes Code desktop client and the yet-to-be-publicly-released flagship version Agnes-2.5-Pro, which has already completed internal training.

After reading the paragraph above, you may still be confused — what AI? What Flash? Haven't heard of it.

However, the word "free" carries real weight. Free means unlimited trial and error, unlimited iteration — no need to scribble classical Chinese prompts to save tokens, nor does it hurt your wallet at month-end. A free model is a solid choice whether you're an experienced player restocking tokens or a beginner exploring Agent workflows.

That said, if the model's capabilities are too poor — unable to understand requests, requiring repeated revisions — it would be maddening even if free. Fortunately, by capability, Agnes is already approaching the global top tier.

I. Agnes Model Overview

1.1 What is AI?

Agnes AI is the model brand under Singapore-based AI company Sapiens Technology Pte. Ltd. The company was founded in 2024 (some sources say July 2025), with founder and CEO Bruce Yang — an alumnus of Raffles Institution Singapore and NUS PhD candidate in AI. The core team members come from universities including NUS, NTU, MIT, Stanford, UC Berkeley, and UT Austin.

Sapiens AI's financing status:

Round

Time

Amount

Lead Investor

Seed

2025

Undisclosed

—

Series A

February 13, 2026

$10 million

LOOK FORWARD VCC

Cumulative Financing

—

Approximately $20 million

—

The company's positioning is remarkably clear: train models themselves, do not rely on third-party APIs. It is not a model aggregation platform like OpenRouter, but rather self-developed from underlying architecture to application layer. This is precisely why it can offer "unlimited free" access — the cost structure is entirely within its own control.

As of June 2026, Agnes AI has over 6 million users, processing more than 5.41 trillion tokens per week (3.25 trillion text tokens, 2.16 trillion multimodal tokens). What does this number mean? Roughly equivalent to freely outputting millions of novels' worth of text every week.

1.2 What is Flash?

Agnes AI's model lineup revolves around three categories:

Text / Agent Models

Model

Release Date

Positioning

Status

Early

Fast conversation, content generation, simple multimodal tasks

✅ Released

Before May 2026

Coding, reasoning, Tool Calling, Agent workflows, image understanding

✅ Current flagship

Next-gen coding & Agent flagship model, fully upgraded

✅ Released

Expected late July 2026

Flagship general reasoning, benchmarking against Claude Opus 4.8 / GLM-5.2

🚧 Pro Preview

Image Models

Model

Purpose

Text-to-image, image-to-image, fast generation

High-precision image-to-image, image editing

Video Models

Model

Purpose

Text-to-video, image-to-video, keyframe animation

Additionally, there is the creative platform Pavo (web-based AI创作 studio) and the desktop IDE tool Agnes Code, unified account system, one credit system covering all capabilities.

1.3 Core Specifications of Agnes-2.5-Flash

Based on official documentation and release information, the key parameters of Agnes-2.5-Flash are as follows:

Dimension

Value

Context Window

256K tokens

Max Output

65K tokens

Multimodal Input

Text + Image URL

Thinking Mode

Compatible (deep thinking activated for coding/debugging scenarios)

Tool Calling

Compatible (Function Calling)

Stream Output

Compatible

API Compatible

OpenAI SDK + Anthropic thinking field compatible

Pricing

Compared to the previous generation agnes-2.0-flash, the official emphasis highlights significant improvements across the following dimensions:

1.4 Compatibility with OpenAI / Anthropic

Clearly compatible. Agnes AI provides an OpenAI-compatible API:

Also compatible with Anthropic-style thinking fields. This means if your project is already using OpenAI or Anthropic SDKs, you only need to replace the base_url and model name to switch — integration cost is minimal.

1.5 Free?

Many will ask: To what extent is it free? Are there hidden limitations?

The actual situation is:

Item

Description

Cost

Input and output both $0

Validity

Unlimited (officially committed)

RPM Limit

Free tier: 30 RPM (actually about 20 RPM executable)

Model Selection

Can only use Agnes AI's own models

Third-party models

Requires platforms such as Agnes Code, charged by membership (pricing not yet announced)

In other words: All of Agnes AI's own models are free, covering text, images, and video. If you use third-party platforms like OpenRouter to access them indirectly, that's a different matter. RPM is sufficient for daily development, but if you need to run large-scale automated Agents, you may need to upgrade to a Token Plan or enterprise plan.

II. Technical Innovation Highlights

In the previous chapter, the reader learned about Agnes's "hard metrics": free, free API, free multimodal. But what truly makes this approach work is not marketing speak, but a complete technical stack from underlying training to upper-level scheduling.

Below, starting from the simplest intuition, we will step by step break down what unique technical optimizations Agnes has that others do not.

2.1 Reinforcement Learning's "Anti-Collapse" Algorithm: DSPO

First, answer a fundamental question: How does a model get smarter?

Large model training has several steps: first, pre-training on massive text (reading extensively), then fine-tuning with human feedback (learning conventions), and finally reinforcement learning (RL) — letting the model learn optimal strategies through "trial and error + rewards/punishments" in complex tasks.

But RL has a fatal weakness: reward signals in complex tasks are too sparse, and frequent failures can cause policy collapse.

For example, asking an AI to write a research report involves roughly: understand the requirement → search information → analyze and organize → write conclusions. If you only give a "reward upon completion" signal at the end, the model has no idea which step was done well and which was wrong. Worse, in search scenarios, most attempts fail, reward signals are nearly zero, gradients become unstable, and training collapses directly.

This is the challenge that mainstream algorithms like DeepSeek's GRPO encounter — they perform excellently on simple reasoning tasks, but easily "learn to collapse" in Agent tasks requiring multi-round search, tool calling, and long-chain planning.

Agnes's approach is called DSPO (Dynamic-filter Sequence-level Policy Optimization). It doesn't invent a new RL algorithm from scratch, but cleverly combines two well-known ideas.

First, the optimization unit mismatch problem. GRPO optimizes at the token level, but rewards are given for the entire sequence — equivalent to scoring each word individually while judging by the quality of the whole sentence. It's like evaluating a marathon by each step. DSPO switches to sequence-level optimization: instead of calculating gradients for each token separately, it computes a comprehensive correction factor for the entire trajectory. This greatly reduces gradient variance and prevents training collapse.

Second, the sparse reward wastes samples. During Search Agent training, many prompt trial groups are either all successful or all failed — their advantage signals are zero or non-existent, wasting compute. DSPO introduces a dynamic filtering mechanism: only retaining trial groups with "both success and failure" for training, ensuring each batch has non-zero, valid advantage signals.

An analogy makes it easier to understand:

This is not a minor improvement. According to Sapiens AI's paper on DSPO published on arXiv, the Agent trained on a 7B model improved by 34.1% compared to the previous baseline model (same-size GRPO), and surpassed a 14B model by nearly 9% relatively — a model with only half the parameters defeated a larger opponent.

What is the perceptible value to users? The model is more stable, Agent task reliability is higher, and search and reasoning are no longer "gambling" but "stable output."

2.2 CodeAgents: Using "Pseudocode" Instead of Natural Language for Agent Collaboration

Having discussed how to strengthen a single model, now let's talk about a different challenge: how can multiple AI Agents communicate with high quality? Currently, using multi-Agent decomposition for complex tasks with division of labor has various detailed implementation approaches, but Agnes builds on this to develop new optimizations.

Suppose you have a research Agent platform with three roles:

If you use natural language to let these three Agents collaborate, what happens? The Planner says "help me look up Southeast Asian tech investment trends," the Solver searches up a bunch of information, and then the Replanner reviews the results and finds them wrong, asking the Planner to adjust direction again. The problem is: everyone's language understanding is ambiguous. "Look up" can mean a deep report or a three-sentence summary. "Tech investment" can mean venture capital, M&A, government funds... the more ambiguity, the more rework.

The core idea of CodeAgents is: don't use natural language as the communication protocol between Agents, but instead use structured pseudocode.

Specifically, how?

Encode the interactions of Planner, Solver, and Replanner into a structure like this:

This is not real Python code, but a typed pseudocode format:

This is like the API convention in software engineering: frontend and backend communicate via JSON schema rather than sending each other voice messages on WeChat.

This formatted approach has some similarity to MCP. MCP is a protocol standard for AI applications to connect external tools and data sources (e.g., how to call search engines, databases, APIs), while CodeAgents is an orchestration framework internal to the Agent platform — it governs how roles like Planner, Solver, and Replanner exchange information, not how to integrate external tools.

In other words, MCP addresses "how AI uses tools," while CodeAgents addresses "how multiple AIs collaborate with high quality." In actual models, the two can coexist: internal multi-Agent communication follows CodeAgents-style structured orchestration, while external tool calling can use MCP for integration.

What does this bring?

Metric

Compared to Natural Language Baseline

Input tokens reduced

Output tokens reduced

Planning accuracy improvement

Absolute value +3 to +36 percentage points

VirtualHome success rate

56% (new SOTA)

Data source: CodeAgents paper on arXiv — Benchmark link: 2507.03254.

Note that this differs from the earlier rough claim of "40-70% token savings" — the paper's original data range is broader, with the actual figures being **input-side savings of 55-87%, output-side savings of 41-70%**.

Of course, this also means that with the same model capability, Agnes can serve more users without proportionally increasing inference costs. This is the real math behind "free": not relying on subsidies, but on saved tokens.

2.3 Universal Verifier: Scoring "Subjective Questions"

Bruce Yang himself calls it the "secret sauce," not open-sourced.

High benchmark scores ≠ good actual experience. You can rank first on math problems but be terrible at writing PPTs, producing research reports, or even "whether to speak in a group chat" — such subjective tasks. Agnes's core model scenarios are precisely these kinds of tasks with no standard answers.

For example:

The traditional approach is to find human annotators to score these subjective tasks — low efficiency and not scalable.

Agnes's approach is Universal Verifier (UV): an internal evaluation platform that can score and provide reward signals for subjective tasks.

According to Bruce Yang's own words: "Otherwise, the model won't know which direction to improve. It doesn't even know where to go."

UV's working method is roughly two layers:

First layer: Evaluator during training. For subjective tasks, UV generates scores and feedback for the output, telling the model "what was done well, what needs improvement." It doesn't just check whether the final answer is correct, but also whether the reasoning path is reasonable.

Second layer: Continuous monitoring after launch. UV tracks real-world feedback signals in the background, and if it detects a drop in response rate for a certain task type, it triggers retraining.

The traditional research → release → launch → discover problems cycle is too slow. UV turns it into a continuous closed loop:

Yang's exact words: "UV is my secret sauce — I haven't open-sourced it." It has no public details, but from its capability positioning, it is the key bridge connecting benchmarks and real user experience.

2.4 The "Middle Layer" of the Generation Platform: Intent Translation, Not Agent Innovation

Frankly speaking, **"three Agents collaborating" is not a new concept.** From GAN's Generator vs Discriminator, to RLHF's Response vs Reward Model, to various frameworks' Solver/Reviewer/Replanner division of labor, the industry has long done this.

What needs to be clarified here is not "who used three Agents," but a more specific challenge:

When the user's input is vague natural language, but the downstream generation model (image/video Diffusion) requires precise structured instructions, what goes in the middle?

Agnes's approach is to add an Intent Understanding Agent between the user and the Generation/Judge Agent. The core logic is simple:

So the true value of this "three-Agent" design lies not in the number of agents, but in the LLM serving as an intent translation layer, converting the ambiguity of natural language into deterministic instructions that the generation model can execute. It doesn't address the "how to organize multiple Agents" novelty, but rather the engineering challenge of "how to connect text generation models and image/video generation models in the same workflow."

In one sentence: This is not a new innovation, but a sound engineering architecture for a specific scenario (multimodal generation). The only thing worth recording is that it was explicitly named and modelized.

Multi-Agent development is not fresh, but the ability to standardize the use of this idea without relying on users to manually harness it — indeed without requiring users to do anything — and to guarantee stability and effectiveness, to optimize to this degree — that is a truly noteworthy innovation, not mere self-packaging.

2.5 Proprietary Model Routing: Not All Problems Deserve the Largest Model

Agnes is not just one model, but a model family. From 7B to 30B+ parameter scales, there are specialized small model versions for different tasks.

Yang's statement is clear:

"I don't need to outperform GPT-4 on every dimension, I just need to exceed it in specific categories — research, PPT generation, group chat assistance. Each mode in my platform has a dedicated model."

The routing logic is as follows:

User Request

Possible Route To

"What's the weather tomorrow?"

Small, fast model

"Generate PPT"

PPT-specialized model

"Image editing"

Image 2.0/2.1 Flash

"Text-to-video"

Video V2.0

"Complex code debugging + Agent workflow"

Agnes-2.5-Flash / Pro

This is essentially multi-model task routing, similar to the "router" concept in MoE, but Agnes's routing granularity is coarser — it doesn't pick experts within a model, but selects调度 targets across models.

Direct impact on cost: approximately half of user traffic is routed to small or specialized models, with only a small fraction of requests consuming maximum compute resources. This means Agnes's cost structure is not dominated by "using the flagship model for every call," but by on-demand allocation. This is one of the economic foundations of "free."

2.6 Real User Data Flywheel

The final piece: the source of data.

Agnes AI has 6 million users, with over 200,000 daily active users. This means a massive amount of real, diverse, and even non-mainstream language and minority-language usage data is generated daily.

The training flow is roughly:

This is a positive cycle:

Why can Agnes achieve this flywheel? Because aggregation platforms (such as OpenRouter) only forward requests and cannot see the complete interaction closed loop between users and models. Agnes trains its own models, provides its own API, and accumulates its own data — this closed loop is complete.

2.7 Speed as Product: Must Complete Within One Minute

Running through all the above techniques is the final principle: speed is the model itself.

Yang's hard standard is: users can complete a full research task within one minute. Why one minute? Psychological research shows that users' perception of waiting is non-linear:

Waiting Duration

Perception

Within 3 seconds

"Instant"

Within 10 seconds

"A bit slow"

30 seconds

Attention starts to drift

60 seconds

Most patience exhausted

To achieve this speed, every technique above must work in coordination: proprietary models reduce inference costs, DSPO improves training stability (reducing trial and error), CodeAgents saves tokens, the intent translation layer reduces generation rework, model routing avoids "using a sledgehammer to crack a nut," and the data flywheel continuously improves effectiveness.

In one sentence: **"Speed as Product" is not an engineering goal optimized after model launch, but a model philosophy that guides the entire technical stack design from the very beginning.**

2.8 Technical Moat Summary Table

Technical Dimension

Core Innovation

Comparison with Traditional Approaches

Paper/Evidence Source

Fully self-developed Text/Vision/Video, no aggregation

Aggregation vs. Self-production

Coruzant

Sequence-level optimization + dynamic filtering to prevent collapse

Token-level GRPO is unstable on Agent tasks

arXiv:2510.09255

Pseudocode-structured multi-Agent communication

Natural language inter-Agent communication is redundant and ambiguous

arXiv:2507.03254

Internal scoring and continuous evaluation for subjective tasks

Relying solely on benchmarks cannot measure real quality

Coruzant

LLM translates vague user needs into precise instructions for generation models

Multi-Agent architecture is not new; value lies in bridging natural language → Diffusion

TechLead Journal #246

7B–30B+ distributed by task, half traffic on small models

Fixed use of one flagship model for all requests

AI Magazine

Real user data → pre-training → distillation → DSPO

Static dataset training

Sapiens AI Case Study

Complete research within one minute

Quality first, then speed up

Coruzant

III. Performance Cross-Evaluation: Who Can a Free Model Compete With?

Just saying "free" is meaningless; what people care about most is: how strong is it really?

3.1 Claw-Eval: The Touchstone of Agent Capability

Claw-Eval is one of the most reference-worthy Agent benchmarks today. It evaluates whether AI can autonomously plan, call tools, and execute multi-step tasks in real environments, like a real software engineer.

Rank

Model

Pass@3

1

Claude Opus 4.6

70.4%

2

Step 3.7 Flash

68.3%

3

Claude Sonnet 4.6

67.8%

...

...

...

Agnes-2.0-Flash ranked 9th globally with a 60.9% Pass@3, surpassing Gemini and MiniMax, entering the global top ten. Agnes-2.5-Flash's specific score on this benchmark has not yet been publicly disclosed, but according to the official statement, the 2.5 version comprehensively surpasses the 2.0 version across all benchmarks.

This ranking carries significant weight: Claw-Eval uses the Pass@3 metric from three independent tests — not running once and giving an average, but requiring success across all three attempts. This means the model must not only "occasionally get it right" but be stable and reliable.

3.2 SWE-bench Verified: Real-World Software Engineering Tasks

SWE-bench Verified measures a model's ability to fix real GitHub issues, regarded by the industry as the "golden standard for coding ability."

Officially disclosed data:

Model

SWE-bench Verified

Notes

Agnes 2.5 Pro

~82.7

Internal evaluation, approaching or reaching top-tier level

GPT-5.5

~84.4 / 88.7*

Closed-source flagship

DeepSeek V4 Pro

80.6

Open-source flagship

Hy3

78.0

Open-source MoE

*GPT-5.5 has two versions under different evaluation systems: standard version ~84.4, Codex CLI version ~88.7.

Agnes 2.5 Pro's 82.7 is already very close to GPT-5.5. But note, this is the Pro version's result — the Flash version's absolute ranking has no specific public data from the official. According to LinkedIn and multiple media reports, Flash surpassed GLM 5.2 and DeepSeek 54 Pro on several sub-items across multiple benchmarks, placing it in the same competitive tier as open-source models at the same level.

3.3 Multilingual SWE: Multilingual Coding

Model

Multilingual SWE

Agnes 2.5 Pro

~78.7

Other models

Official only showed relative rankings, no specific scores

3.4 Horizontal Comparison with Major Competitors

Dimension

Agnes-2.5-Flash

DeepSeek V4 Flash

GLM-5.2

Kimi K3

GPT-5.5

Claude Opus 4.8

0.07

0.28

0.64

30

75

256K

128K

128K

1,000,000

1,000,000

❌

✅ Apache 2.0

✅ Apache 2.0

✅ Planned

❌

❌

Top 10

Strong

Strong

Strong

Top tier

Top tier

Top tier

Top tier

Text + Image input

Text-focused

Text + Vision

Text + Image + Video

Text + Image + Video

Text + Image + Video

✅ Permanently free

❌ Paid API

❌ Paid API

❌ Paid API

❌ Paid API

❌ Paid API

Several observations:

3.5 Cost-Effectiveness Calculation

Suppose you write code for 4 hours a day, generating approximately 10,000 tokens of output per hour on average:

Item

Daily Consumption

Monthly Consumption

GPT-5.5 Price

Claude Price

Agnes Price

Output tokens

~240K

~7.2M

$216

$540

Input tokens

~480K

~14.4M

$72

$216

Just this one figure alone can save you considerable money every month with Agnes. Of course, this doesn't mean you should rely entirely on free models, but at least for daily development, prototype validation, and learning exploration, it is fully trustworthy and can be used as a primary model.

IV. Summary: Is It Worth Using?

I mainly use the desktop version of OpenCode,搭配 third-party APIs to use various models, and Agnes is one of them.

Based on my experience in user intent understanding, code generation, workflow optimization, code debugging and problem-solving, document writing, and fact-checking, my evaluation of Agnes 2.5 Flash is: higher than DeepSeek V4 Flash, comparable to or slightly below DeepSeek V4 Pro, not as good as Hy3 but the gap is not large.

Considering the free aspect and stability (I use Nvidia's free DeepSeek V4 Pro, but its quota is small and inherently unstable),综合 the performance, I have personally made Agnes 2.5 Flash my default model, occasionally using DeepSeek V4 Flash for cross-validation.

Suitable Scenarios

Unsuitable Scenarios

In One Sentence

Agnes-2.5-Flash is the only model on the market today that simultaneously achieves "free + top-tier capability + complete multimodal." It may not rank first in every dimension, but in terms of "getting sufficiently good experience for the least money," there is currently no competitor.

Reference Sources

Official Documentation and Release

Academic Papers

Benchmark Rankings

Technical Analysis and Interviews

Benchmark Data News

Epilogue

Agnes AI's story is fascinating. A Singapore startup, $20 million in financing,不做 aggregation platform, training their own full-stack models from text to image to video, and then offering everything free.

This approach is radical by any industry's standards — but the reason they dare to do this is because they have something others don't: a complete proprietary model stack. No need to pay "tolls" to OpenAI, Anthropic, or Google, so they can drive the price to zero.

Free is never the endpoint, but an entry point to let users taste the sweetness first. But if Agnes-2.5-Pro can approach Claude Opus 4.8 as advertised, then this entry point may be far deeper than many imagine.

As for whether it can truly retain users? How long can this model run? I also have no answer, but perhaps the answer lies in our hands.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry