Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo, MiniMax @aiDotEngineer
Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo, MiniMax  @aiDotEngineer
Uploaded October 2026 | Updated November 2026, 3 days ago
A faster model only helps if the finished job gets faster. Morgan Suo of MiniMax uses one real research task to test the changes inside the M3.1 Flash model: an analyst checking whether circuit-board suppliers are turning AI demand into cash. The changes are sparse attention on long inputs, a 4-bit KV cache, low-precision experts, speculative decoding, server load and reasoning budgets. Each one saves compute, and each can quietly cost accuracy or wait time. She then shows how to judge a model switch: check sources and calculations, count the full cost of a usable report, and make the result repeatable.

Chapters:
- 00:00 Intro
- 01:44 The research job
- 03:20 Profit vs. cash flow
- 04:22 What changed inside the model
- 05:09 Sparse attention on long inputs
- 07:53 4-bit KV cache
- 08:59 Low-precision experts
- 10:20 Speculative decoding
- 11:39 Throughput vs. wait time
- 13:09 Reasoning budgets
- 14:41 How to evaluate a model switch
- 16:12 Recap

Morgan Suo, MiniMax: minimax.io

Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo, MiniMaxImagination Engineering: Live in the future and then build whats missing.Build Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAIProving Alpha: How a Hedge Fund Measures AIs ROI - Hari Kumar, Contour Asset ManagementAIE CODE: Google DeepMind, Anthropic, SpaceXSI, TypeSafe, Lovable, Ramp, etc (System One theme)The Cost of Accessibility in AI — Jemima Abu, CAISDont Vibe Code Your Agent Infrastructure - Akshat Bubna, ModalHow to Know When an AI Agent Can Work Unsupervised — Evan Kormos, CoinbaseYou Cant Tell What an Agent Skill Will Do by Reading It - Roey Ben Chaim, ZenityGive Your On-Call Its Own AI Teammate - Swaapnika Motapothula, Chinmay Gandhi & Vlad Tsvang, MetaWhy AI Writing Still Reads Like Slop — Louis-François Bouchard, Towards AIHow a Trading Firm Ships New AI Models the Same Day - Noyan Tokgozoglu, Optiver
AI Engineer |

Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo, MiniMax

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER