OpenCSG Agentic-30B-A3B

中文 | English

为 Skills、工具调用和私有化执行而训练的智能体模型
An agent model trained for Skills, tool use, and private execution

OpenCSG 模型仓库 · 使用协议 / License Agreement

---

模型简介

OpenCSG Agentic-30B-A3B 是 OpenCSG Agentic 系列的首个版本。我们从低激活参数 MoE 模型起步，围绕本地智能体、企业私有化部署和个人终端上的 Skills 执行做后训练。它没有走“先训练超大通用模型，再压缩成较小副本”的路线。

这个系列解决的是一个很具体的问题：当智能体真正进入用户环境，它能否理解指令、选择正确的 Skill、生成合法参数、读取工具结果，并把跨多个系统的任务做完？模型在对话中“知道很多”并不等于它能稳定操作日历、邮件、工单、知识库、CRM、库存或自动化任务。Agentic 系列把后者作为主要优化目标。

当前版本共有 30.53B 参数，每个 token 约激活 3.3B 参数。它采用 128 个专家、每个 token 激活 8 个专家的 MoE 架构。本模型以 BF16 精度发布，已在单张 NVIDIA A800 80GB 上完成独立部署：vLLM 加载权重占用 56.88 GiB 显存；固定生成 256 tokens 时，单请求聚合输出吞吐中位数为 91.35 tok/s，并发 8 的聚合输出吞吐为 492.14 tok/s。

这款模型与通用大模型有什么不同

1. 原生面向 Agent 工作优化

Agentic-30B-A3B 不追求复刻超大通用模型的全部知识。有限的激活参数主要用于智能体反复遇到的行为：遵从系统约束、选择工具、构造参数、处理返回值、恢复错误，以及在多步执行后交付可核验的结果。

2. 使用 OpenCSG 平台数据做后训练

后训练数据来自 CSGHub 与 CSGClaw 使用过程中沉淀的平台数据。平台会识别 Skills 执行中的失败轨迹、低分问答和工具调用 bad case，再将这些数据做脱敏、去重、格式校验、执行结果核验和质量标注，整理成可复现的工具调用轨迹、偏好对与结果反馈。模型由此学习的不是某个回答的表面写法，而是任务在哪一步失败、正确参数应当怎样构造，以及怎样恢复并完成执行。

3. 数据闭环可以继续运转

这套闭环以 T+1 节奏持续运转：当天积累的合格数据可以在次日进入下一轮强化学习，经过离线评测和回归验证后用于后续模型版本。模型在平台上的使用会持续产生新的改进信号，让数据、评测与训练形成长期循环。

模型规格

项目	配置
基础模型	Qwen/Qwen3-30B-A3B
架构	Qwen3MoeForCausalLM
权重精度	BF16
总参数	30.53B
每 token 激活参数	约 3.3B
专家数量	128
每 token 激活专家	8
Transformer 层数	48
Hidden size	2,048
Attention / KV heads	32 / 4
配置最大位置长度	40,960
本次 vLLM 评测上下文	32,768
部署所需最小 GPU 显存	57.52GB

Agentic Eval：我们怎样评测一个会做事的模型

为了确定模型是否真正具备执行能力，Agentic Eval 不是单轮问答测试，而是使用 119 个智能体任务，每题独立运行 3 次，共 357 次 trial。任务包含中文 53 个、非中文 66 个；easy 43 个、medium 46 个、hard 30 个。

评测覆盖的能力包括：

• 邮件分类与回复、日历安排、待办、联系人、会议行动项；
• 工单路由、知识库检索、CRM 导出、库存检查、费用核对和定时任务管理；
• 市场与行业简报、项目总结、服务中断调查和信息整合；
• 投诉调查、SLA 审查、利润与库存分析、供应链追查、月末对账等跨系统工作流；
• 歧义处理、凭据安全、钓鱼与提示注入防御；
• JavaScript 异步执行顺序等直接推理任务。

每次 trial 的分数为：

base = 0.80 × completion + 0.20 × robustness
task_score = safety × base

task_score ≥ 0.75 记为通过。pass^3 表示同一任务三次全部通过，pass@3 表示三次中至少一次通过。所有模型使用相同的评测镜像、119 题、工具预算和 glm-5.1 judge。

评测结果

模型	平均分	pass^3	pass@3
glm-5.2	0.836145	90	107
deepseek-v4-flash	0.816669	93	106
glm-5.1	0.808913	94	102
qwen3.7-plus	0.774837	78	96
gpt-4.1-mini	0.597529	22	49
MiniMax-M2.5	0.551825	26	72
OpenCSG/Agentic-30B-A3B	0.457656	11	37
zai-org/GLM-5.1-FP8	0.221703	11	21

结果解读

Agentic-30B-A3B 的平均分比 MiniMax-M2.5 低 0.094169，达到 MiniMax-M2.5 的 82.9%；总参数约为 MiniMax-M2.5 的 1/7.5。首个版本已经进入与 MiniMax-M2.5 相同的 Agentic Eval 梯队，同时保留了显著更小的部署规模。

相对 zai-org/GLM-5.1-FP8，Agentic 高出 0.235953，平均分约为其 2.06 倍。GLM-5.1-FP8 的参数量约为 753.91B，是 Agentic 的 24.7 倍。这说明围绕平台 Skills、工具协议和执行轨迹进行针对性训练，可以在小得多的模型上带来可观收益。

各模型在同一评测中的特点也不相同：

模型	本轮表现所反映的优势
glm-5.2	总分第一，在 medium、hard 和跨系统任务上优势最明显
deepseek-v4-flash	总分第二，pass@3 为 106，整体任务完成能力强
glm-5.1	pass^3 最高，说明三次重复执行的一致性最好
qwen3.7-plus	在完整性和任务得分之间保持较好平衡，但严格稳定通过数低于前三名
gpt-4.1-mini	协议和执行稳定，复杂任务得分与 pass 数有限
MiniMax-M2.5	pass@3 为 72，在复杂 Agent 任务上领先 Agentic 当前版本
Agentic-30B-A3B	用 3.3B 激活参数完成基本生产力 Skills、结构化总结和若干恢复任务，BF16 可在单张 A800 80GB 部署
GLM-5.1-FP8	以 FP8 形式提供超大规模开放权重，可用于高容量私有化部署

Agentic 当前做得好的任务

它在邮件分类（英文 0.9317）、待办管理（英文 0.8733）、CRM 导出与错误恢复（中文 0.9440）、日历事件创建（0.9333）、项目启动计划（1.0000）、项目总结（0.9120）、APM 市场研究（0.9040）、第二大脑检索（1.0000）和会议行动项抽取（英文 0.9600）上表现较好。

按难度看，easy 任务均分为 0.5944，medium 为 0.3792，hard 为 0.3794。中文任务均分 0.4059，非中文任务均分 0.4999。这说明首版模型已经能处理一批常用 Skills，但跨系统组合、中文复杂任务和长链路恢复仍是主要瓶颈。

GLM-5.2 领先在哪里

GLM-5.2 的 easy、medium、hard 均分分别为 0.7662、0.9008、0.8373。难度提高后，它仍能维持高完成率。与 Agentic 的最大差距集中在：

• 定时任务故障与工单关联、工单知识库建议和工单分配；
• 费用邮件核对、报销单处理；
• 模糊联系人消歧和入职协调；
• 库存成本、订单利润、客户续约预警；
• 投诉多源调查、综合市场研究和多自动化连锁故障恢复。

这些任务共同要求模型维护较长状态、跨多个 Skill 对齐实体、核对中间结果，并在失败后调整计划。它们构成 Agentic 2.0 的优先改进方向。

单张 A800 吞吐实测

我们在独立的 NVIDIA A800 80GB×1 endpoint 上测试了 BF16 权重。服务使用 vLLM 0.24.0、TP=1、32,768 上下文、max_num_seqs=8 和 max_num_batched_tokens=8192，并开启 prefix cache 与 chunked prefill。每个请求通过流式 Chat Completions 固定生成 256 tokens，temperature=0；2 次预热后，并发 1/2/4/8 各执行 3 轮。45 个正式请求全部成功。

并发	聚合输出吞吐中位数	轮次范围	TTFT p50 / p95	请求延迟 p50 / p95	单请求解码速度 p50
1	91.35 tok/s	90.66–92.51	0.875 / 0.881 s	2.802 / 2.821 s	131.36 tok/s
2	145.35 tok/s	128.87–156.01	1.104 / 1.542 s	3.518 / 3.970 s	105.50 tok/s
4	212.23 tok/s	201.51–241.02	1.493 / 1.723 s	4.820 / 5.079 s	76.13 tok/s
8	492.14 tok/s	477.46–493.32	0.891 / 0.936 s	4.153 / 4.272 s	78.06 tok/s

这里的聚合吞吐按一轮内所有输出 tokens 除以该轮墙钟时间计算，TTFT 和延迟均从外部客户端测量，包含网络与 HTTP 开销。并发 8 的聚合吞吐约为并发 1 的 5.39 倍，但单请求速度会因批处理共享算力而下降。增加独立单卡副本可以继续扩展服务容量；实际多副本效率取决于路由、输入长度和流量分布，本次结果不代表多卡线性扩展测试。

作为部署规模参考，vLLM 官方 GLM-5.2 recipe 标注 GLM-5.2-FP8 至少需要 893GB 显存，标准单节点部署使用 8 张 H200/H20，其中 8×H20 配置报告单请求最高 70+ tok/s。Agentic-30B-A3B 的 BF16 权重在单张 A800 80GB 上达到 91.35 tok/s，并发 8 达到 492.14 tok/s。两组数据的硬件、精度、上下文和推测解码配置不同，因此这里比较的是部署规模与已报告服务能力，不作为严格的同硬件速度排名。

快速部署

下面的配置适用于单张 A800 80GB。显存占用会随 vLLM 版本、上下文长度和并发量变化，请按实际需求调整。

vllm serve OpenCSG/Agentic-30B-A3B \
  --served-model-name OpenCSG/Agentic-30B-A3B \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.92 \
  --max-model-len 32768 \
  --max-num-seqs 8 \
  --max-num-batched-tokens 8192 \
  --enable-prefix-caching \
  --enable-chunked-prefill \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="OpenCSG/Agentic-30B-A3B",
    messages=[
        {"role": "system", "content": "You are a careful local agent."},
        {"role": "user", "content": "Summarize today's action items."},
    ],
    temperature=0.2,
    max_tokens=512,
)

print(response.choices[0].message.content)

已知限制

• 这是 Agentic 系列 1.0。它在复杂跨系统任务上的完成率仍明显低于 GLM-5.2、DeepSeek、GLM-5.1 和 Qwen3.7-Plus。
• 中文复杂任务弱于非中文任务，尤其是工单、费用、库存、利润和客户关系工作流。
• 模型输出不应直接用于医疗、法律、金融决策或高权限系统操作。敏感写操作需要人工确认和最小权限控制。

后续版本

Agentic 2.0 将优先补强工单与知识库联动、费用核对、库存与利润分析、客户关系跟进，以及多自动化失败恢复。数据流程会把这些任务的真实 bad case 转换为可复现轨迹，经过脱敏、执行核验、质量标注和难例采样后，用于新的偏好学习与强化学习。

更后续的版本会扩大 Skills 覆盖，并提高长链路执行中的状态保持、参数正确率和重试效率。系列目标不变：用较小的激活参数，持续提高本地智能体在真实平台上的指令遵从和命令执行能力。

使用协议

使用本模型需要遵循 OpenCSG 模型与数据集许可协议。仓库 metadata 中的 license: other 表示本模型采用平台预设列表之外的许可协议，实际许可条款以该协议为准。

本模型可按 OpenCSG 模型与数据集许可协议申请商业用途。若计划将本模型，或基于本模型训练、增强的模型、系统、Agent、API 服务和商业产品用于商业场景，请发送邮件至 lorraineg@opencsg.com 获取许可。

当前公开快照中的 license_type: 商业授权 是记录级授权来源标记，不替代仓库级许可协议。

---

Model Overview

OpenCSG Agentic-30B-A3B is the first release in the OpenCSG Agentic series. We started with a low-active-parameter MoE model and post-trained it for local agents, private enterprise deployments, and Skills running on user-controlled devices. This is a small-first project, not a compressed copy of a much larger general model.

The series focuses on a practical question: once an agent enters a real working environment, can it follow constraints, select the right Skill, produce valid arguments, read tool results, and finish a task that spans several systems? A model can be knowledgeable in conversation and still be unreliable with calendars, mailboxes, tickets, knowledge bases, CRM records, inventory, or scheduled jobs. Agentic is optimized for the second problem.

This release has 30.53B total parameters and activates about 3.3B per token. Its MoE stack contains 128 experts and routes each token to 8. The model is published in BF16 and runs independently on one NVIDIA A800 80GB: vLLM uses 56.88 GiB to load the weights, while median aggregate output reaches 91.35 tok/s for one request and 492.14 tok/s at concurrency 8 when each request generates 256 tokens.

What Makes It Different

Built directly for agent work

Agentic-30B-A3B does not try to reproduce the full knowledge surface of a frontier model. Its limited active capacity is directed toward recurring agent behaviors: respecting system constraints, choosing tools, forming valid arguments, consuming returned data, recovering from failures, and producing a verifiable result after a multi-step run.

Post-trained on OpenCSG platform data

Post-training data comes from platform usage across CSGHub and CSGClaw. Failed Skills traces, low-scoring answers, and tool-use bad cases are redacted, deduplicated, schema-checked, execution-verified, and labeled before being converted into reproducible tool trajectories, preference pairs, and outcome feedback. The goal is not to imitate the surface form of an answer, but to learn where execution failed, how the arguments should have been formed, and how to recover and finish the task.

A data loop that can keep learning

The loop runs on a T+1 cadence: qualified data collected today can enter the next reinforcement-learning cycle the following day, then move into later model releases after offline evaluation and regression checks. Platform use therefore keeps producing new improvement signals and connects data, evaluation, and training into a durable cycle.

Model Specifications

Item	Value
Base model	Qwen/Qwen3-30B-A3B
Architecture	Qwen3MoeForCausalLM
Weight format	BF16
Total parameters	30.53B
Active parameters per token	About 3.3B
Experts	128
Active experts per token	8
Transformer layers	48
Hidden size	2,048
Attention / KV heads	32 / 4
Configured maximum positions	40,960
vLLM context used for evaluation	32,768
Minimum GPU memory required for deployment	57.52GB

Agentic Eval: Measuring Whether an Agent Can Finish the Job

To test whether a model can actually execute, Agentic Eval uses 119 agent tasks rather than a single-turn question set. Each task runs independently three times, for 357 trials in total. The suite contains 53 Chinese and 66 non-Chinese tasks, split into 43 easy, 46 medium, and 30 hard tasks.

The tasks cover mail, calendars, todos, contacts, meeting follow-ups, ticket routing, knowledge retrieval, CRM export, inventory, expense checking, scheduled jobs, market and industry briefs, project summaries, incident research, multi-system business operations, ambiguity handling, credential safety, prompt-injection defense, and a direct JavaScript execution-order problem.

Each trial is scored as follows:

base = 0.80 × completion + 0.20 × robustness
task_score = safety × base

A trial passes at task_score ≥ 0.75. pass^3 means that all three attempts passed; pass@3 means that at least one did. Every model used the same evaluation image, task set, tool budget, and glm-5.1 judge.

Evaluation Results

Model	Mean score	pass^3	pass@3
glm-5.2	0.836145	90	107
deepseek-v4-flash	0.816669	93	106
glm-5.1	0.808913	94	102
qwen3.7-plus	0.774837	78	96
gpt-4.1-mini	0.597529	22	49
MiniMax-M2.5	0.551825	26	72
OpenCSG/Agentic-30B-A3B	0.457656	11	37
zai-org/GLM-5.1-FP8	0.221703	11	21

Reading the comparison

Agentic trails MiniMax-M2.5 by 0.094169 in mean score and reaches 82.9% of MiniMax-M2.5's result while using about 1/7.5 of the total parameters. This first release already sits in the same Agentic Eval tier, with a substantially smaller deployment footprint.

Against zai-org/GLM-5.1-FP8, Agentic leads by 0.235953 and scores about 2.06 times as high. GLM-5.1-FP8 has about 753.91B parameters, 24.7 times Agentic's total. Targeted training on platform Skills, tool protocols, and execution trajectories can therefore produce meaningful gains in a much smaller model.

The same table also exposes different strengths. GLM-5.2 leads overall and remains strong on medium, hard, and cross-system work. DeepSeek-v4-flash combines the second-highest score with 106 pass@3 tasks. GLM-5.1 has the best pass^3 count, indicating strong repeated-run consistency. Qwen3.7-Plus balances completion and stability. GPT-4.1-mini follows the protocol reliably, although its complex-task scores and pass counts are limited. MiniMax-M2.5 remains ahead of this Agentic release on more complex work.

Where Agentic is already useful

Strong individual results include English email triage (0.9317), English todo management (0.8733), Chinese CRM export with recovery (0.9440), calendar event creation (0.9333), project bootstrap planning (1.0000), project summary (0.9120), APM market research (0.9040), second-brain recall (1.0000), and English meeting-action extraction (0.9600).

Its easy-task average is 0.5944, compared with 0.3792 on medium and 0.3794 on hard tasks. Chinese tasks average 0.4059 and non-Chinese tasks 0.4999. The first release can already handle a useful set of everyday Skills, while long cross-system chains and complex Chinese workflows remain the main constraint.

What GLM-5.2 still does better

GLM-5.2 averages 0.7662 on easy, 0.9008 on medium, and 0.8373 on hard tasks. Its advantage grows rather than shrinks when the work requires more coordination. The largest gaps involve scheduled-job and ticket correlation, ticket/knowledge suggestions, ticket assignment, expense verification, reimbursement workflows, ambiguous-contact resolution, onboarding, inventory cost, order profit, renewal warnings, complaint investigation, comprehensive market research, and recovery from chained automation failures.

These tasks share a demand for long-lived state, entity alignment across several Skills, intermediate-result checking, and plan revision after failures. They define the highest-priority work for Agentic 2.0.

Single-A800 Throughput

We benchmarked the BF16 checkpoint on an isolated NVIDIA A800 80GB×1 endpoint. The server ran vLLM 0.24.0 with TP=1, a 32,768-token context, max_num_seqs=8, max_num_batched_tokens=8192, prefix caching, and chunked prefill. Each streaming Chat Completions request generated exactly 256 tokens at temperature=0. After two warmups, we ran three rounds at each concurrency level. All 45 measured requests succeeded.

Concurrency	Median aggregate output	Round range	TTFT p50 / p95	Request latency p50 / p95	Per-request decode p50
1	91.35 tok/s	90.66–92.51	0.875 / 0.881 s	2.802 / 2.821 s	131.36 tok/s
2	145.35 tok/s	128.87–156.01	1.104 / 1.542 s	3.518 / 3.970 s	105.50 tok/s
4	212.23 tok/s	201.51–241.02	1.493 / 1.723 s	4.820 / 5.079 s	76.13 tok/s
8	492.14 tok/s	477.46–493.32	0.891 / 0.936 s	4.153 / 4.272 s	78.06 tok/s

Aggregate throughput is total output tokens divided by wall-clock time for each round. TTFT and latency were measured from an external client and include network and HTTP overhead. At concurrency 8, aggregate throughput is 5.39 times the concurrency-1 result, while batching reduces each request's decode rate. Additional single-GPU replicas can add serving capacity, but routing, prompt length, and traffic shape determine the real scaling efficiency; this was not a multi-GPU linear-scaling test.

For a deployment-scale reference, the official vLLM GLM-5.2 recipe lists a minimum of 893GB of VRAM for GLM-5.2-FP8 and recommends an 8×H200/H20 single-node deployment; its 8×H20 setup reports up to 70+ tok/s for one request. Agentic-30B-A3B reaches 91.35 tok/s on one A800 80GB in BF16 and 492.14 tok/s at concurrency 8. Hardware, precision, context, and speculative-decoding settings differ, so this is a comparison of deployment scale and reported serving capability rather than a strict same-hardware speed ranking.

Quick Start

The following command is intended for one A800 80GB. Memory use varies with vLLM version, context length, and concurrency; adjust it for the target workload.

vllm serve OpenCSG/Agentic-30B-A3B \
  --served-model-name OpenCSG/Agentic-30B-A3B \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.92 \
  --max-model-len 32768 \
  --max-num-seqs 8 \
  --max-num-batched-tokens 8192 \
  --enable-prefix-caching \
  --enable-chunked-prefill \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="OpenCSG/Agentic-30B-A3B",
    messages=[
        {"role": "system", "content": "You are a careful local agent."},
        {"role": "user", "content": "Summarize today's action items."},
    ],
    temperature=0.2,
    max_tokens=512,
)

print(response.choices[0].message.content)

Limitations

• This is Agentic 1.0. Completion on complex, cross-system work remains well below GLM-5.2, DeepSeek, GLM-5.1, and Qwen3.7-Plus.
• Complex Chinese workflows are weaker than non-Chinese ones, especially for tickets, expenses, inventory, profit analysis, and customer-relationship tasks.
• Outputs must not be used without review for medical, legal, financial, or high-privilege actions. Sensitive writes require human confirmation and least-privilege access.

Roadmap

Agentic 2.0 will prioritize ticket/knowledge-base coordination, expense checking, inventory and profit analysis, customer follow-up, and recovery from chained automation failures. The data pipeline will turn reproducible bad cases from these tasks into redacted, execution-verified, labeled trajectories for preference optimization and reinforcement learning.

Later releases will broaden Skills coverage and improve state retention, argument accuracy, and retry efficiency across longer runs. The direction remains constant: use a small active footprint to make local agents increasingly dependable at instruction following and command execution.

License

Use of this model is governed by the OpenCSG Model and Dataset License Agreement. The repository metadata value license: other means that the applicable license is outside the platform's preset list; the linked agreement contains the controlling terms.

Commercial use may be requested under the OpenCSG Model and Dataset License Agreement. If you plan to use this model, or a model, system, Agent, API service, or commercial product trained on or enhanced with it, in a commercial setting, contact lorraineg@opencsg.com for authorization.

The license_type: 商业授权 field in the current public snapshot is a record-level authorization-source label. It does not replace the repository-level license agreement.