Jun 12
Didn't save the URL of a model page in @MyAirTrade? No problem, search with "ModelName myairtrade" and the 1st result will get you to model page where all their available assets are listed! #AircraftForSale #AircraftForLease #EnginesForSale #EnginesForLease
ALT MyAirTrade
17
Mar 21
co-authored by opencode
<modelname>
would go a long way
idt its tasteless in the long run but yeah, when only one does it...
5
716
Mar 17
🚨큰거 온다🚨
테슬라 코리아 웹사이트 코드에서 YL 모델 변형에 대한 단서 발견 $TSLA
코드 발견
테슬라 코리아의 컨피규레이터 코드(웹팩 번들의 index.js)에 새로운 트림 검사 항목이 추가되었습니다.
JavaScriptconst isE80 = TrimName === ' Y L';
enhancedData.TrimName = isE80 ? TrimName : new
String (TrimName).replace (`${ modelName} `, ");
E80은 YL 모델(아마도 장거리 주행용 변형 모델)의 내부 코드로 취급되고 있습니다
1
1
2
168
Mar 12
Sometimes when I’m bored I write my prompts as if I was a Victorian era businessman. Starting with “Dear modelname” and ending with faithfully yours, official name, title and all that.
It always sends me when a model gets the joke and sends me a formal letter as a reply abck
4
925
Mar 1
这个仓库是一个“视觉驱动 Android 自动化代理”原型：通过 ADB 抓屏 Qwen3-VL 识别界面 ADB 执行动作，实现用自然语言驱动手机操作。整体设计清晰，目标明确，适合做端侧/半端侧 GUI Agent 的实验和二次开发。
从结构看，核心只有 4 个 Python 文件：
• "phoneagent.py"：任务主循环与 ADB 执行器，负责截图、调用视觉模型、执行 tap/swipe/type/wait/terminate。
• "qwenvlagent.py"：Qwen3-VL 推理与 tool-call 解析，把模型输出转换为内部动作字典。
• "ui.py"：Gradio 控制台，包含任务面板、日志面板、设置面板（分辨率、温度、token 等）。
• "qwenvlutils.py"：图像/视频输入整理。
另外 "config.json" 负责运行参数（分辨率、重试、步间延迟等）。
执行链路是典型 Agent Loop：
1. "adb shell screencap" 抓屏并 pull 到本地。
2. 把图片和任务上下文喂给 Qwen3-VL。
3. 解析 "<toolcall>"，映射到内部动作（click->tap、坐标归一化）。
4. 用 "adb shell input ..." 执行动作。
5. 循环直到模型返回 "terminate" 或达到上限。
这套流程可读性强，便于替换模型和动作后端。
仓库的主要优点：
• 功能闭环完整：CLI 与 Web UI 都可直接跑。
• 上下文机制简洁：保留最近动作历史，足够支撑短任务。
• 对显存和输入尺寸有一定工程化处理（图像缩放、可选 Flash Attention）。
• 分辨率支持自动校验/修正，降低“点击偏移”的常见问题。
当前实现里也有一些明显风险与可改进点：
• 模型配置不一致：README 说默认 Dense 4B/8B，但代码默认值在多处仍是 30B-A3B；且 "PhoneAgent" 初始化 "QwenVLAgent" 时没有把 "config" 中 "modelname" 传进去，导致配置项名义存在但实际不生效。
• 滑动动作精度损失：模型给了起止坐标，但执行层只保留方向（up/down/left/right）并固定从屏幕中心滑动，复杂界面上容易失败。
• "ui.py" 的“停止任务”只改了全局标志位，执行主循环未消费该标志，停止按钮大概率无法立即中断实际任务。
• 重试逻辑语义偏差："maxretries" 与总 cycle 计数耦合，不是“单步失败重试次数”，在长任务中可能提前或异常终止。
• ADB 命令执行采用 "shell=True" 拼接字符串，若未来把未清洗输入透传进命令，存在命令注入风险。
• 缺少自动化测试、requirements 锁定与基准任务集，不利于回归验证。
适用场景：
• 研究/演示型 GUI Agent。
• 内网或实验环境下的 Android 自动化脚本编排。
• 作为更复杂 Phone Agent 的最小可运行骨架。
如果要把它推进到“更可靠的生产级自动化”，建议优先做这几件事：
1. 打通配置一致性（统一默认模型、让 "model_name" 真正生效）。
2. 保留并执行真实 swipe 起止坐标，不只用方向模板。
3. 重构中断与重试机制（可中断循环、区分 step retry 与 task cycle）。
4. 给动作解析和 ADB 执行层补单元测试，给典型任务补端到端回归。
5. 明确依赖版本并提供一键安装脚本（requirements/lockfile）。
整体评价：这是一个“思路正确、骨架清楚、可快速上手”的仓库，作为 Phone GUI Agent 的学习与原型基础很合适；但在可靠性、可维护性和工程安全性上，还需要一轮系统性加固。
github.com/OminousIndustries…
6
22
2,878
Feb 24
echo str_replace("claude","DeepSeek",$modelName);
şunu da mı ekleyememişler...
2
89
Feb 19
Update on my Go-agent-sdk (v.0.3.0) !!!
It now supports multiple LLM providers out of the box
One interface, swap one line, your agent works with OpenAI, Anthropic, Gemini, or any OpenAI-compatible endpoint (Groq, DeepSeek, Cerebras, Mistral, and more)
What's new:
- ChatProvider interface - every provider implements CreateChat ModelName, agent doesn't care which one you pick
- Native Anthropic provider - full translation layer for Claude's Messages API including tool calling
- Native Gemini provider - full translation layer for Google's generateContent API including the tool call detection quirk (Gemini returns STOP even for tool calls)
- 12 OpenAI-compatible base URLs built in - Groq, Cerebras, DeepSeek, Together, Fireworks, Mistral, Moonshot, DashScope (Qwen), ZAI (GLM), Anyscale
- Still zero dependencies. Pure standard library Go.
Each provider is a self-contained translator. Adding a new one means writing one file.
Nothing else changes. Remains fast concurrent and clean
Providers available now - @GeminiApp , @AnthropicAI , @OpenAI , @Kimi_Moonshot ,@deepseek_ai , @Alibaba_Qwen ,@cerebras , @Zai_org and so much more !
You can take a look at the repo here -
github.com/parthshr370/Go-ag…
Demo below showing how easy it is to swap the Providers !
x.com/parthshr370/status/201…
Feb 1
For the last few weeks I have been trying to get my hands dirty with Golang.
And the best way to do it was to build something that I was confident in but purely written in Golang.
Here is my v0.1: a Go-agent-sdk -- github.com/parthshr370/Go-ag…
Inspired by all your agent frameworks but written with minimal dependencies - just the standard library (net/http, encoding/json, reflect). Zero external dependencies.
While building and playing around with Go I realised that it being fast and concurrent-native can be a goldmine for future agent work. The lack of such SDKs made it even more compelling to build it on my own.
So from the message handling, API requests, JSON handling, and even the tool calling engine, all built from scratch from first principles.
Some credit where needed that helped me understand and make it even better:
- @pydantic AI's very first minimal commit gave me the foundational structure to build on
- @CamelAIOrg 's modular agent architecture and role-playing patterns informed my design approach
- @Kimi_Moonshot K2.5 helped refine the documentation to match Go's conventions—turning my informal comments into proper Go docstrings that feel native to the ecosystem
- @openrouter for the backend (with plans to add more providers soon)
5
11
601
Laravel tip.
Laravel's factory magic methods like `for()` eliminate boilerplate in your tests.
Define BelongsTo relationships once, and Laravel automatically generates convenient `->for{ModelName}()` methods for intuitive test data creation.
1
6
45
2,095
Jan 20
So you want to find a video, EG: Connecting Orca slicer to the Falshforge [modelname]
1
2
16
7 Dec 2025
ただ、↑↑のプラグインそのままのコマンドだと、ツクールMZ内の変数を受け付けないので…。
計算式を使うなら、スクリプトで入力する必要があるのです。
意味は知らなくていいからイベントコマンドの『スクリプト…』に以下を丸写しして、『変数番号』だけ書き換えれ。。。#ゲーム制作 #ゲーム開発
//2Dモデル表示 (変数＝モデル名)
const args = {modelName:$gameVariables.value(変数番号)};
PluginManager.callCommand(this,"Live2DInterfaceMZ","show", args);
//2Dモデルサイズ変更 (変数＝モデル名、拡大率)
const args = {modelName:$gameVariables.value(変数番号),scale:$gameVariables.value(変数番号)};
PluginManager.callCommand(this,"Live2DInterfaceMZ","scale", args);
//2Dモデル移動 (変数番号＝モデル名、X、 Y、 時間)
const args = {
modelName:$gameVariables.value(変数番号),
x:$gameVariables.value(変数番号),
y:$gameVariables.value(変数番号),
dur:$gameVariables.value(変数番号)};
PluginManager.callCommand(this,"Live2DInterfaceMZ","pos", args);
//2Dモーション再生 (変数番号＝モデル名、モーション名、ループするかtrue/false)
const args = {modelName:$gameVariables.value(変数番号),
motionName:$gameVariables.value(変数番号),loop:"false"};
PluginManager.callCommand(this,"Live2DInterfaceMZ","motion", args);
//2D表情変更 （変数番号＝モデル名、表情名）
const args = {modelName:$gameVariables.value(変数番号),expressionName:$gameVariables.value(変数番号)};
PluginManager.callCommand(this,"Live2DInterfaceMZ","expression", args);
//2Dモデル消去 （変数番号＝モデル名
const args = {modelName:$gameVariables.value(変数番号)};
PluginManager.callCommand(this,"Live2DInterfaceMZ","hide", args);
6
321
20 Nov 2025
На роутерах помимо мака светится ProductName или DeviceID. У айфонов оно выглядит как iPhone (John's). У андроидов <modelname> John's.
Умный человек может изменить его вручную, но не станет угонять тачку у бывшего работодателя.
2
11
598
17 Oct 2025
What are JavaScript Objects?
In JavaScript, objects allow you to store and organize data in a structured way. Objects are collections of key-value pairs, where each key is a string, symbol or number and each value can be of any data type, including other objects, functions, and primitive values.
For example, think of a car. It’s an object, with the properties like color, model and manufacture year.
// object
const car = {
color: "black",
modelName: "Fortuner",
yearOfManufacture: "2021",
};
Here, the term car represents an object. The keys are color, modelName, and yearOfManufacture, and their corresponding values are "black", "Fortuner", and "2021".
To put it simply, a JavaScript object is like a container that lets you store and organize properties and methods related to something.
Objects are used for representing complex data structures, organizing code, and creating classes and instances through object-oriented programming.
1
4
72
6 Aug 2025
For prompts:
They generally gravitate to similar solutions. But, as an example, a team who included something like "ALWAYS use a timeout when calling long running commands" will probably seem better than a competitor's agent that doesn't provide that guidance. Lots of finesse here.
Tools:
Anthropic (Claude Code) probably doesn't call a use_gemini_model(model_name) or use_openai_model(modelname) tool despite competitor models being better in certain situations.
Accessibility:
How easy is it to connect to my agent. Can I connect from anywhere. Can it run unsupervised in a remote environment. Can I attach to it at anytime. Are the outputs readable and coherent; what data is exposed to the user. This is somewhat subjective and depends on the use case.
Real World Example:
Cursor and Copilot run the same underlying models in an editor built on top of VS Code, but there's a difference.
1
2
81
25 Jul 2025
The “S“ after the modelname indicates it’s actually a show model which has the retail look on purpose. (which makes it even more special!)
1
239
8,329
23 Jul 2025
are you having trouble with it?
for nvidia just do
VLLM_USE_V1=0 vllm serve /root/location/or/qwen/modelname --tensor-parallel-size 4 --dtype bfloat16 --cpu-offload-gb 0 --gpu-memory-utilization 0.98 --port 9900 --enable-prefix-caching --quantization fp8 --trust-remote-code --host 0.0.0.0 --api-key "sk_boomer" --disable-log-requests --served-model-name "gpt4" --enable-expert-parallel --enable-auto-tool-choice --tool-call-parser qwen3_coder
You'll need to install the nightly for this though
uv pip install vllm --pre --extra-index-url wheels.vllm.ai/nightly/
2
4
438