---
title: "minhleduc (@firefrog) | Hashnode"
description: "Just another Dev guy. Read the latest articles by minhleduc on Hashnode."
url: "https://hashnode.com/@firefrog"
---

@firefrog

# minhleduc

@firefrog · Ho Chi Minh City, Vietnam · Joined July 2026

Just another Dev guy.

### [Why Production Agents Are Ditching Pure RAG for Context Caching](https://minhleduc.hashnode.dev/why-production-agents-are-ditching-pure-rag-for-context-caching?utm_source=hashnode&utm_medium=feed)

Three months ago, our agentic workflow hit a cost wall. We were running a multi-step research agent that performed iterative document analysis. Every user request triggered between 12 and 18 agent ste

8 min read · [minhleduc.hashnode.dev](https://minhleduc.hashnode.dev/why-production-agents-are-ditching-pure-rag-for-context-caching?utm_source=hashnode&utm_medium=feed)

[Aug 16](https://hashnode.com/posts/why-production-agents-are-ditching-pure-rag-for-context-caching/6a818b738e94ec1956f398d0)

### [Kimi K3 has 2.8T parameters. Moving them was the hard part.](https://minhleduc.hashnode.dev/kimi-k3-has-2-8t-parameters-moving-them-was-the-hard-part?utm_source=hashnode&utm_medium=feed)

Inside Stable LatentMoE and Attention Residuals — the two changes that made a 3T-class model trainable.

15 min read · [minhleduc.hashnode.dev](https://minhleduc.hashnode.dev/kimi-k3-has-2-8t-parameters-moving-them-was-the-hard-part?utm_source=hashnode&utm_medium=feed)

[Aug 8](https://hashnode.com/posts/kimi-k3-has-2-8t-parameters-moving-them-was-the-hard-part/6a769f693e4681a62cc1b73a)

### [How Kimi K3 reads a million tokens](https://minhleduc.hashnode.dev/how-kimi-k3-reads-a-million-tokens?utm_source=hashnode&utm_medium=feed)

The road from softmax to Kimi Delta Attention — and why deleting positional encoding was worth 5.5 points.

12 min read · [minhleduc.hashnode.dev](https://minhleduc.hashnode.dev/how-kimi-k3-reads-a-million-tokens?utm_source=hashnode&utm_medium=feed)

[Aug 8](https://hashnode.com/posts/how-kimi-k3-reads-a-million-tokens/6a769f4eefa957e947735e73)

### [5 Tiny Language Models for Tool Calling (Part 3)](https://minhleduc.hashnode.dev/5-tiny-language-models-for-tool-calling-part-3?utm_source=hashnode&utm_medium=feed)

The benchmark nobody quotes when they tell you small models are the future. NVIDIA Research published a position paper called "Small Language Models are the Future of Agentic AI". I believe the argume

12 min read · [minhleduc.hashnode.dev](https://minhleduc.hashnode.dev/5-tiny-language-models-for-tool-calling-part-3?utm_source=hashnode&utm_medium=feed)

[Aug 5](https://hashnode.com/posts/5-tiny-language-models-for-tool-calling-part-3/6a72db5132b016de0e5d59d2)

### [5 Tiny Language Models You Need to Know in 2026 (Part 2)](https://minhleduc.hashnode.dev/5-tiny-language-models-you-need-to-know-in-2026-part-2?utm_source=hashnode&utm_medium=feed)

Gemma 4 E2B, Qwen3.5-4B, SmolLM3-3B, Phi-4-mini and Ministral 3 3B — the five small language models worth knowing in 2026, and when to pick each.Gemma 4 E2B, Qwen3.5-4B, SmolLM3-3B, Phi-4-mini and Ministral 3 3B — the five small language models worth knowing in 2026, and when to pick each.

10 min read · [minhleduc.hashnode.dev](https://minhleduc.hashnode.dev/5-tiny-language-models-you-need-to-know-in-2026-part-2?utm_source=hashnode&utm_medium=feed)

[Aug 5](https://hashnode.com/posts/5-tiny-language-models-you-need-to-know-in-2026-part-2/6a72d74a22919e3b0a50214f)
