DEV Community: threerouter

nvidia isn't just selling the picks and shovels anymore

threerouter — Wed, 12 Aug 2026 02:11:50 +0000

it's mining now.

Nemotron 3.5 Lightning dropped: 30B param MoE, 3B active, Apache 2.0, runs agent loops up to 4x faster, 1M token context. ships with NeMo Switchyard, an open router that farms long tasks to the small model and saves the frontier stuff for planning.

day-0 on SGLang, OpenRouter, Fireworks. free for commercial use.

the open tier keeps getting better without asking permission. the "frontier gap" is starting to look like a pricing gap.

AI #OpenWeights

qwen理应牛逼起来

threerouter — Tue, 11 Aug 2026 09:49:43 +0000

阿里又整活了！Qwen-MM-Plugins 发布。
把多模态塞进插件，直接给现有 Agent 装上“眼睛”和“手”——读图、剪视频、搞 3D/CAD 全搞定。
从“能看”到“能干”，这才是真·多模态 Agent。
Alibaba just dropped another banger! Qwen-MM-Plugins is here.
They stuffed multimodal capabilities into a plugin—instantly giving any Agent framework eyes and hands. It reads, edits, and processes visuals, video, and 3D/CAD.
Moving from "seeing" to "doing." This is the real Multimodal Agent.

全世界让你token自由的这几家你可以收藏

threerouter — Fri, 07 Aug 2026 10:55:20 +0000

全世界把 DeepSeek V4 Flash 卖的比官方还便宜 6 倍的总共有三家

Command Code https://commandcode.ai/docs/plans/go

OpenCode https://opencode.ai/go

threerouter https://www.threerouter.com/

这个价格的模型套餐使用体验是这样的：用不完，根本用不完

DeepseeP-v4-flash真强，加上codex无敌了

threerouter — Fri, 07 Aug 2026 04:27:39 +0000

Codex 现在能让 DeepSeek 直接干活了。

这个开源项目把 deepseek-v4-flash 做成 Codex 原生子 Agent：

Codex 负责规划决策，DeepSeek 负责写代码、搜索、重构、测试……

不是取代，是给 Codex 配个超强助手。

成本更低，上下文更省，一个人就能搭出自己的 AI 开发小团队

Using DeepSeek V4-Flash—super affordable.

threerouter — Tue, 04 Aug 2026 12:53:11 +0000

Well. DeepSeek V4-Flash: 8 trillion tokens in a single day. most-called model on earth.

openai cut Luna 80% last week. still ~60% pricier than this thing. wild.

anyone actually running prod workloads on it?

Generate an SVG Use deepseek-v4-flash

threerouter — Tue, 28 Jul 2026 04:11:22 +0000

Prompt: Generate an SVG of a pelican riding a bicycle . Use deepseek-v4-flash what do you think?

Open source is the future Open source is the future. The countdown is on for open-weight Kimi-K3—a top-tier contender poised to hit the arena. With unbeatable cost-efficiency, open LLMs aren't just viable; they're untouchable.

threerouter — Mon, 27 Jul 2026 11:27:06 +0000

My 3-Month Hermes Agents Journey: Lessons Learned (and Disagreements)

threerouter — Mon, 27 Jul 2026 06:48:10 +0000

I've been using Hermes Agents for 3 months now, and I see a lot of great advice floating around. But after real production use, here's my honest take — what I agree with, and what I respectfully push back on.

What I Fully Agree With

Context is still king. Period. Prompt engineering → context engineering → graph engineering — the buzzwords change, but the core remains the same. Your soul.md, user.md, agent.md matter, but so do your company structure, codebase, and knowledge directories. Without rich context, you're just chatting with a parrot.

*2. Optimize your .md files relentlessly. They bloat over months. Keep them simple and executable. Otherwise, you're literally burning tokens on noise. Less is more.

Optimize your Skills frequently. Hermes auto-generates skills faster than you can blink. They're "progressive disclosure" in theory, but in practice they become wordy novels. Simplify. Make them executable. Your token bill will thank you.
CRON + hooks is NOT enough for multi-agent collaboration. Research agent → product agent → engineering agent → that's a workflow, not a trigger chain. I had to build custom action queues to handle state and dependencies properly. This is where the real engineering begins.
Observability and auditability are foundational. When your research agent hits arXiv or does web searches, require it to save intermediate artifacts (PDFs, HTML, markdown) AND metadata. Without this, you're flying blind. Debugging becomes impossible.
Eval is the ceiling of your agent application. Write eval standards first — quantitative metrics (script-based, fast, stable) and qualitative metrics (LLM-based). Don't let LLMs judge your quantitative metrics. Scripts are quicker and more reliable.

What I Partially Agree With

JSON and Markdown can get messy. Yes, using SQLite for structured data and Markdown for unstructured data makes sense. JSON as intermediate artifacts for observability? Great. But saying "Markdown/JSON instructions don't work" misses the point — the issue is usually prompt design, not the format itself. Don't blame the format for bad prompting.

What I Respectfully Disagree With

Don't use Hermes memory.md or other memory providers. Hard disagree. Memory is what makes LLM applications intelligent over time, not just clever in the moment. The problem isn't memory — it's how you use it. If your agent isn't proactively calling memory, fix the trigger strategy, don't abandon memory entirely. Without memory, your agent has amnesia. That's not a feature.
Don't use Hermes profile for multi-agent applications. Also disagree. Yes, profiles aren't hierarchical. Yes, sharing common values across roles is tricky. But "there's no better way" doesn't mean we abandon the best available tool. Use base templates + override patterns. Maintain them well. Profile management is a discipline, not a design flaw.

My Bottom Line
After 3 months of real-world Hermes usage, here's what I've learned:

Memory and profiles aren't the problem — lazy engineering is.Context, eval, observability, and workflow orchestration are where you should spend your energy.And for heaven's sake, optimize your tokens. Your wallet will notice.

What's been YOUR experience with Hermes Agents?

Are you using memory providers successfully?

How are you handling multi-agent workflows?
What's your eval strategy?

Drop your thoughts below — I'd love to learn from your journey too.

HermesAgents #AI #LLM #AgentEngineering #MultiAgent #AIDevelopment #PromptEngineering #TechLessons

Like this? Share it with your network and let's start a conversation about what actually works in production agent systems.

open models are the future

threerouter — Mon, 27 Jul 2026 02:01:09 +0000

I agree：
Jensen Huang and Elon Musk back open models—and open models are the future.

Mira Murati's Thinking Machines released Inkling under Apache 2.0: 975B total, 41B active, 1M context, multimodal. Pitch from the team: a fine-tunable open base, not a leaderboard chaser.

DeepSeek and Kimi set the bar. This is a real answer.

Heads up — / Quick heads-up:DeepSeek just killed the old endpoints." Still calling DeepSeek through deepseek-chat or deepseek-reasoner? They go offline for good at 15:59 UTC today, with no fallback. Update your model string or your requests start failin

threerouter — Fri, 24 Jul 2026 11:15:35 +0000

GPT-5.6 Sol Ultra has proven the decades-old Cycle Double Cover Conjecture

threerouter — Sat, 11 Jul 2026 11:13:37 +0000

GPT-5.6 Sol Ultra has proven the decades-old Cycle Double Cover Conjecture, showing stunning capability. Still, I care more about costs. If a single proof costs millions in tokens, it is merely academic spectacle and useless for most developers.

GPT-5.6 Sol Ultra is coming to Codex, Expect a huge leap in coding capability. Staytuned!

threerouter — Mon, 06 Jul 2026 07:36:04 +0000

GPT-5.6 for Codex? Awesome — except for the price tag.