AMD acquires Taalas: model-specific AI silicon
AMD agreed to buy Taalas on 6 August 2026, eight months after Nvidia licensed Groq for a reported $20B. Both deals say inference is splitting from training.
Edward Xi Yang
Ertas Blog
Field notes on building custom AI models: fine-tuning, on-device inference, and the cost maths behind owning your own models.
All 594 articles, grouped by topic.LFM2.5 sizes compared: 230M vs 350M vs 1.2B on-device
Liquid AI ships LFM2.5 at three sizes. Benchmarks, export sizes, and what we learned fine-tuning two of them into shipped products, including the failure no metric caught.
Edward Xi Yang
Fine-Tuning for Indie Devs: Inside the $10 Ertas Lite Plan
The cheapest way to train a custom AI model you own. What $10/mo unlocks for hobbyists and indie devs, which models you can train, and where it runs out.
Edward Xi Yang
Chatty Valley: an on-device AI mod for Stardew Valley (Part 1)
How I fine-tuned a 1.2B on-device AI model to give Linus real dialogue inside Stardew Valley, running on your CPU with no API key. Part 1 of the Chatty Valley build log.
Edward Xi Yang
Gemma 4 E2B and FunctionGemma 270M: Tool Calling
Google's FunctionGemma (270M) and Gemma 4 E2B (2B) are the smallest credible function-calling models of 2026, built as bases you fine-tune and ship.
Edward Xi Yang
Llama Stack on a Phone: Self-Hosted Llama Agents With a Fine-Tuned Llama 4 Model
Meta's Llama Stack is the canonical reference architecture for Llama-based agents. Combine it with a fine-tuned Llama 4 derivative and the Swift/Kotlin client SDKs and you get a complete agent stack running entirely on the user's phone.
Edward Xi Yang
Mastra + Vercel AI SDK + On-Device GGUF: A TypeScript Mobile Agent Stack With No API Costs
TypeScript-first mobile builders don't have to use Python agent frameworks. Mastra and the Vercel AI SDK plus a fine-tuned 4B model running on-device through llama.cpp produce a complete agent stack with zero per-token costs.
Edward Xi Yang
Phi-4-Mini vs Gemma 4 vs Qwen3-4B: Tool Calling
Phi-4-Mini, Gemma 4 E4B and Qwen3-4B on BFCL v4 tool calling, phone latency and post fine-tune accuracy, plus a four-question tree for picking your base.
Edward Xi Yang
Replacing OpenAI in OpenAI Agents SDK With Your Fine-Tuned Local Model
The OpenAI Agents SDK is intentionally model-agnostic. Swap the OpenAI client for an Ertas-trained model running on Ollama and you keep the developer experience while killing per-token costs. A drop-in tutorial.
Edward Xi Yang
Pydantic AI On-Device: Fine-Tune Qwen3-4B for Type-Safe Mobile Agents
Pydantic AI brings type safety and FastAPI ergonomics to LLM agents. Combine it with a fine-tuned 4B model running on-device via llama.cpp and you get production-grade agents in mobile apps with zero API costs and validated outputs by construction.
Edward Xi Yang
Pydantic AI vs LangGraph for Fine-Tuned Agents
Pydantic AI gives you type-safe linear agents, LangGraph durable stateful graphs. A 2026 decision matrix, plus why fine-tuning makes either one reliable.
Edward Xi Yang
Nous Hermes Agent vs Hermes 4: The Difference
Nous Research ships two things called Hermes: an open-weight model family from 2025 and a self-improving agent framework from 2026. Here is which to use when.
Edward Xi Yang
The 2026 Open Source AI Model Landscape
A comprehensive snapshot of the open-weight AI model ecosystem as of April 2026 — Chinese-lab dominance, MoE architectural defaults, the unified thinking-mode pattern, and what it all means for production deployments.
Edward Xi Yang
Why Chinese Labs Now Dominate Open-Source AI
By April 2026, Chinese labs hold the top five open-weight models on aggregate intelligence benchmarks. The pattern isn't an accident — it reflects strategic, structural, and economic differences between US and Chinese AI development that took years to play out.
Edward Xi Yang
The Effective Context Length Problem: Why 1M Tokens Isn't Really 1M Tokens
Models advertised with 1M or 10M token context windows don't actually retain useful retrieval accuracy across that full range. Here's what 'effective context' really means, why it matters for production deployments, and how to design around the gap.
Edward Xi Yang
Mixture of Experts (MoE) Architecture in 2026
Why MoE is the default for flagship open-weight models in 2026, how the design evolved from Mixtral to DeepSeek V4, and what it means for deployment.
Edward Xi Yang
A/B Testing Cloud API vs On-Device AI in Production
How to run a fair A/B test between your cloud API and on-device model in a live mobile app. Metrics, cohort design, statistical significance, and the metrics that actually matter.
Edward Xi Yang
How to Add AI to Your Mobile App: A Developer's Decision Guide
A comprehensive guide covering every approach to adding AI features to iOS and Android apps. Cloud APIs, on-device models, and hybrid architectures compared with real cost and performance data.
Edward Xi Yang
AI in Android Apps: ML Kit, Cloud APIs, and On-Device LLMs Compared
Three paths to AI in your Android app. Google ML Kit for common tasks, cloud APIs for full LLM capability, and on-device models via llama.cpp for cost and privacy. A practical comparison for Kotlin developers.
Edward Xi Yang
Your AI API Bill Will 10x When Your App Gets Users
The cost math most AI tutorials skip. Your API bill scales linearly with every user, and the real multipliers are worse than the pricing page suggests. Here's what happens at 1K, 10K, and 100K MAU.
Edward Xi Yang
AI API Pricing for Mobile: The Real Cost Per User
How to calculate the true cost of AI per mobile app user. Provider comparison, hidden multipliers, and the unit economics that determine whether your AI feature is sustainable.
Edward Xi Yang
AI API Rate Limits Will Throttle Your Mobile App at Scale
Rate limits from OpenAI, Anthropic, and Google are designed for controlled usage, not mobile apps with thousands of concurrent users. Here is where the limits hit and what happens when they do.
Edward Xi Yang
Why Your AI App Feels Slow: Mobile API Latency
AI API calls add 500-3,000ms of latency to every interaction. On mobile that is the gap between a feature users love and one they abandon. Here is the fix.
Edward Xi Yang
AI Features Mobile Users Actually Want (2026)
Research-backed list of AI features that drive retention and engagement in mobile apps. What users want, what they ignore, and how to prioritize AI features based on actual behavior data.
Edward Xi Yang
Deploy custom AI models, no ML expertise required.
Free plan, no card. Paid plans from $10/mo USD.