Author

Manjunath Janardhan

@manjunath.shiva

articles: 9
total claps: 2,219
top claps: 732
avg read: 10.2 min
paywalled: 9
active: 2025-12-03 → 2026-07-17

Wiki topics

LLM · Large Language Models · 6 AGT · AI Agents · 4 💻 · Programming · 2 OPS · LLMOps & Inference · 2 🔓 · Open Source · 2 🏃 · Running & Endurance · 2 🎬 · Film & Television · 2 🥊 · Combat Sports · 1

Top tags

#apple-silicon ·5 #mlx ·4 #turbo-quant ·3 #qwen ·2 #speculative-decoding ·2 #claude-code ·2 #local-llm ·2 #quantization ·1 #clawdbot ·1 #ai-assistant ·1

MA Manjunath Janardhan · Data Science Collective

How I run 122B-parameter LLMs on a MacBook — outperforming MXFP4 and standard quantization on Apple…

I extended TurboQuant to Mixture-of-Experts models. A 120B LLM that no existing format could fit on 64GB now runs at 44 tokens per second…

Apr 06, 2026 · 7 min read · 732 claps paywalled #turbo-quant #apple-silicon #mlx

How I run 122B-parameter LLMs on a MacBook — outperforming MXFP4 and standard quantization on Apple…

AI TCH SPT

MA Manjunath Janardhan · Data Science Collective

I Turned Two Macs Into an 80B AI Cluster for Free — Exo Is the Open-Source Tool You’ve Been Waiting…

Running Qwen3-Next-80B at 70–80 tokens/second on a home cluster with zero cloud costs

Mar 14, 2026 · 9 min read · 595 claps paywalled #exo #ai-cluster

I Turned Two Macs Into an 80B AI Cluster for Free — Exo Is the Open-Source Tool You’ve Been Waiting…

AI SOC TCH SPT

MA Manjunath Janardhan · AI Advances

FemtoClaw: I Built the World’s Smallest AI Assistant on a $4 Chip

A full AI agent with memory, tools, and Telegram — running on bare metal with 120KB of RAM. No Linux. No Python. Just C.

Mar 01, 2026 · 7 min read · 396 claps paywalled #clawdbot #ai-assistant #esp32

FemtoClaw: I Built the World’s Smallest AI Assistant on a $4 Chip

MDA AI

MA Manjunath Janardhan · Data Science Collective

Run Qwen3–235B on a 16 GB Mac mini: 1.58-bit Ternary Quantization with TurboQuant-MLX

Per-expert disk streaming runs Qwen3–235B-A22B — a flagship open-weight MoE — on the cheapest Mac Apple sells. The headline build is a…

Jul 02, 2026 · 23 min read · 277 claps paywalled #local-llm #mlx #turbo-quant

Run Qwen3–235B on a 16 GB Mac mini: 1.58-bit Ternary Quantization with TurboQuant-MLX

AI TCH

MA Manjunath Janardhan · Data Science Collective

Run Claude Code Locally on a Mac: 65 tok/s with a 4-bit Qwen3.6–27B and DFlash Speculative Decoding

A step-by-step guide to driving Claude Code entirely offline on Apple Silicon — no cloud, no API bill, ~19 GB of RAM, and real-time decode…

Jun 18, 2026 · 11 min read · 81 claps paywalled #claude-code #local-llm #apple-silicon

Run Claude Code Locally on a Mac: 65 tok/s with a 4-bit Qwen3.6–27B and DFlash Speculative Decoding

MDA AI

MA Manjunath Janardhan · Data Science Collective

A Qwen 3.5 122B LLM on a 16 GB Mac mini: MoE Expert Streaming with TurboQuant-MLX

Per-expert disk streaming runs a 122-billion-parameter Mixture-of-Experts model — 3× bigger than RAM — on the cheapest Mac Apple sells…

May 25, 2026 · 12 min read · 60 claps paywalled #turbo-quant #mlx #qwen

A Qwen 3.5 122B LLM on a 16 GB Mac mini: MoE Expert Streaming with TurboQuant-MLX

AI SPT

MA Manjunath Janardhan · Mac O’Clock

LM Studio Bionic on a 64 GB M4 Max: A Fully Local AI Agent with Gemma 4 26B at 103 tok/s

LM Studio’s new agent app read my lecture notes and Jupyter notebook, wrote python-pptx code, fixed its own bugs, and delivered a…

Jul 17, 2026 · 7 min read · 50 claps paywalled #lm-studio #gemma-4 #local-agent

LM Studio Bionic on a 64 GB M4 Max: A Fully Local AI Agent with Gemma 4 26B at 103 tok/s

AI TCH

MA Manjunath Janardhan · Data Science Collective

Building a Local Agentic Browser With Fara-7B, LMStudio, and Magentic-UI (Step-by-Step Guide)

A fully local, privacy-first, agentic system.

Dec 03, 2025 · 6 min read · 26 claps paywalled #agents #agentic-browser #fara

Building a Local Agentic Browser With Fara-7B, LMStudio, and Magentic-UI (Step-by-Step Guide)

AI TCH

MA Manjunath Janardhan · Data Science Collective

Sparse MoE vs Speculative Decoding: The Fastest Way to Run a Local Coding LLM on Apple Silicon

I benchmarked three ways to run a local agentic coding assistant on a Mac. The mixture-of-experts model reached the first token up to 6×…

Jun 19, 2026 · 9 min read · 2 claps paywalled #speculative-decoding #moe #apple-silicon