Author

Old Zhang's AI Learning

AI practitioner specializing in large-model evaluation and on-premise deployment, agents, AI programming, Vibe Coding, general AI, and broader tech trends, with daily original technical articles.

345

Articles

Likes

5.8k

Views

Comments

Latest from Old Zhang's AI Learning

100 recent articles max

Old Zhang's AI Learning

Oct 7, 2026 · Artificial Intelligence

AI-Generated Inference Engine Beats vLLM by 90% at 1,792 tokens/s

Baseten engineers used Claude Code to auto-generate a custom inference engine (VibeQwen) for Qwen-3.6-35B-A3B on B200, achieving 1,792 tokens/s decode (90% faster than vLLM), 12ms TTFT, and 10,307 tokens/s throughput at concurrency 32, via MetaInfer's LLM-as-Compiler loop with accuracy oracle and reusable optimizations, also speeding up SAM 3.1 by 50% in two days.

AI-generated codeB200 GPUClaude Code

0 likes · 11 min read

AI-Generated Inference Engine Beats vLLM by 90% at 1,792 tokens/s

Old Zhang's AI Learning

Oct 7, 2026 · Artificial Intelligence

Perplexity's pplx-decider-v1.1-27b: Open-Weight Decision Model Beats Jev, Runs Locally

Perplexity releases pplx-decider-v1.1-27b, an open-weight decision model based on Qwen3.8-27B that tops the Decision Index benchmark, outperforming Jev and other clones, with non-causal attention and a calibrated decision head for classification, routing, and scoring tasks, deployable on 80GB GPUs.

Decision IndexOpenAI Decisions APIPerplexity

0 likes · 19 min read

Perplexity's pplx-decider-v1.1-27b: Open-Weight Decision Model Beats Jev, Runs Locally

Old Zhang's AI Learning

Oct 6, 2026 · Industry Insights

Codex Declines: SemiAnalysis Reveals Claude Opus 5.5 Delivers 5x Value Over OpenAI

The article analyzes OpenAI's Codex decline, citing a SemiAnalysis report showing Anthropic's Claude Opus 5.5 provides over 5x more API value per dollar than OpenAI, with benchmarks comparing GPT-6.1 Sol, Gemini 3.8 Flash, and Claude Opus 5.5 on throughput and task completion, explaining how slow models disrupt developer flow.

AI coding assistantsClaude Opus 5.5Codex

0 likes · 11 min read

Codex Declines: SemiAnalysis Reveals Claude Opus 5.5 Delivers 5x Value Over OpenAI

Old Zhang's AI Learning

Oct 1, 2026 · Artificial Intelligence

Gemini 4 Argon: 1M Output Tokens, Real-World Engineering Feats & Security Breakthroughs

Google's Gemini 4 Argon introduces 1M output tokens enabling end-to-end industrial engineering agents, demonstrated by 300TiB–1PiB memory savings in data centers, 800k-line Rust migration of Zircon microkernel with 2.7x speedup, and top scores across software engineering, finance, legal, and cybersecurity benchmarks.

1M output tokensGemini 4 ArgonLarge Language Models

0 likes · 17 min read

Gemini 4 Argon: 1M Output Tokens, Real-World Engineering Feats & Security Breakthroughs

Old Zhang's AI Learning

Sep 29, 2026 · Artificial Intelligence

Open-Source macOS Audio Studio for Qwen-Audio TTS Built with Gemini 3.8 Flash

The author details the development of an open-source macOS native application for Alibaba's Qwen-Audio-3.1-TTS-Next model, featuring an AI scriptwriter, native creation studio with voice parameter controls, instant playback, project library with A/B testing, and 42 templates, all built using SwiftUI and Gemini 3.8 Flash.

AI audio generationGemini 3.8 FlashOpen Source

0 likes · 9 min read

Open-Source macOS Audio Studio for Qwen-Audio TTS Built with Gemini 3.8 Flash

Old Zhang's AI Learning

Sep 28, 2026 · Industry Insights

ChatGPT Desktop Overhaul: Chat vs Work Modes, Cloud Sandboxes & 100GB Library

This article reviews ChatGPT's desktop app major update, detailing six key features: Chat vs Work modes with GPT-6 Pro access, cloud VM-backed sandboxes (4-9 cores), 100GB Library with Google Drive integration, image generation parity with web, expanding plugin ecosystem, and Map feature limitations for Chinese users, concluding with the author's transformed workflow.

AI productivityChatGPTGPT-6 Pro

0 likes · 10 min read

ChatGPT Desktop Overhaul: Chat vs Work Modes, Cloud Sandboxes & 100GB Library

Old Zhang's AI Learning

Sep 25, 2026 · Artificial Intelligence

Ternary Bonsai 2: Qwen3.8-27B Compressed to 5.9GB at 98.2% Performance

PrismML's Ternary Bonsai 2 27B uses rotated weight basis and FP16 group-wise scaling to ternary-quantize Qwen3.8-27B to 1.76 bits (5.9GB), achieving 98.2% benchmark retention across coding, reasoning, and agent tasks, with custom CUDA/MLX kernels enabling fast inference on consumer GPUs and Apple Silicon.

Bonsai 2GGUFLLM deployment

0 likes · 15 min read

Ternary Bonsai 2: Qwen3.8-27B Compressed to 5.9GB at 98.2% Performance

Old Zhang's AI Learning

Sep 24, 2026 · Artificial Intelligence

Alibaba's Qoder Unleashes Free Qwen3.8-Flash: Hands-On Test Exceeds Expectations

The author tests Alibaba's Qoder AI coding tool with the newly free Qwen3.8-Flash model, using it to rebuild a complex audio studio app with 14 requirements, demonstrating SubAgent parallelism, site templates, and long-context handling, concluding the free model suffices for most development tasks.

AI coding assistantAI-assisted developmentAlibaba

0 likes · 9 min read

Alibaba's Qoder Unleashes Free Qwen3.8-Flash: Hands-On Test Exceeds Expectations

Old Zhang's AI Learning

Sep 23, 2026 · Artificial Intelligence

Hands-On Test: Qwen-Audio-3.1-TTS-Next Powers Open-Source One-Person Audio Studio

The author tests Alibaba's new Qwen-Audio-3.1-TTS-Next model across diverse scenarios — suspense drama, game NPCs, ads, podcasts — demonstrating its end-to-end generation of layered soundscapes, and releases open-source tools Qwen Audio Studio and z-qwen-audio-studio for local audio production.

AI audio generationQwen-Audio-3.1-TTS-Nextaudio production

0 likes · 14 min read

Hands-On Test: Qwen-Audio-3.1-TTS-Next Powers Open-Source One-Person Audio Studio

Old Zhang's AI Learning

Sep 22, 2026 · Artificial Intelligence

Doubao Work Financial AI Agent: Automating A-Share Equity Research from Data to Contrarian Analysis

The author walks through a full equity research workflow on A-share optical communication stocks using Doubao Work's financial data plugin and built-in skills, demonstrating automated hotspot analysis, conditional stock screening, research report digestion, multi-stock financial comparison with formula-driven Excel output, daily monitoring, and a contrarian risk review via the Stock Analyst work partner, highlighting cross-device collaboration and source-cited rigor.

A-share analysisAI financial agentDoubao Work

0 likes · 14 min read