AI UNDERDOGSDAILY PICK
AI UNDERDOGS
8B 小模型，53% → 99%
8B model: 53% → 99%
antoinezambelli/forge
一个 8B 小模型做 agent 任务
裸跑只有 53% 成功率
加上结构化验证和重试机制之后
直接干到 99%
An 8B model hits 53% raw on agentic
tasks. Add structured validation and retry guardrails —
suddenly it's 99%
小模型的 agent 框架
Agent framework for small models
小模型做工具调用经常瞎编
大多数人直接换 GPT-4 砸钱解决
Forge 的思路是反过来——不换模型
给它套上结构化校验和自动重试
让消费级显卡上跑的模型也能胜任
Small models hallucinate on tool calls — most
devs just upgrade and eat the cost. Forge
flips that: wrap the model in schema validation
and auto-retry so consumer-GPU models can actually handle
it
★ SIGNAL 1
99% 不是玄学
99% isn't magic
Forge 的核心不是推理优化
是工具调用的结构化校验——模型输出的 JSON 不对就重试
工作流每一步都有 schema 兜底
整个框架自带代理服务
所有东西跑在本地，数据不出你的机器
The trick isn't inference tricks — it's schema-level
validation on tool calls. JSON malformed?
Retry. Every workflow step has a schema safety
net. Built-in model proxy, everything runs locally, zero
data leaves your machine
★ SIGNAL 2
162 个 fork 在用
162 forks — people ship with it
67 个 commit
162 个 fork——这不是 demo 放完就跑的项目
最近的提交还在修代理模型发现和 OpenAI 兼容客户端的推理捕获
说明真有人在生产里用
67 commits, 162 forks — this isn't demo-and-abandon
Recent fixes hit proxy model discovery and reasoning
capture in the OpenAI-compatible client. People are actually
shipping this
作者自己说：「Guardrails 把 8B 模型在 agent 任务上的表现从 53% 提到了 99%」
没有吹什么颠覆性突破
就是一个很实在的工程优化——但效果摆在那里
The author puts it plainly: 'Guardrails take an
8B model from 53% to 99% on agentic
tasks.' No grandiose claims — just a solid
engineering layer. The numbers speak
AI UNDERDOGS
自己跑一个试试
Go try it yourself
antoinezambelli/forge
关注 · 每天发现更多 AI 神作
github.com/antoinezambelli/forge