ModelsLongCat 2.0: Meituan's 1.6T MIT-Licensed Model for Agents
Meituan's LongCat 2.0 pairs 48B active parameters with MIT weights and a 1M-token design. What is new in it, how its scores read, and how it sizes up to DeepSeek V4 Pro.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
Picking between models — and testing them on your own work.
Models
ModelsLyria 3 Clip Preview: 30-Second Music Generation at $0.04 a Clip
Lyria 3 Clip generates fixed 30-second music clips through the Gemini API. How it is billed, what a fixed length is good for, and how to prompt and evaluate it.
Read
ModelsLyria 3 Pro Preview: Full-Length AI Songs and What to Build On
Lyria 3 Pro generates full-length songs for $0.08 each. What it controls, what it refuses, and whether to build on the preview now that Lyria 3.5 is GA.
ReadModelsMagnum v4 72B: A Community Fine-Tune Built for Prose
Magnum v4 72B is Anthracite's Qwen2.5 72B fine-tune aimed at Claude 3 style prose. What it is trained on, the two-layer licence, and how to judge its writing fairly.
Read
ModelsMancer Weaver (alpha): A 2023 Roleplay Model in 2026
Weaver is an 8K-context roleplay model from 2023 that is still listed today. What is known about it, what the small window costs you, and when to move off it.
ReadModelsMercury 2 Guide: How a Diffusion Reasoning Model Gets Its Speed
Inception Mercury 2 writes by refining whole drafts in parallel rather than one token at a time. What that buys, what the speed claims measure, and where it fits.
ReadModelsMercury 2.5 Guide: What Changed in Inception's Diffusion LLM
Mercury 2.5 doubles the context to 260K and claims a 40% quality gain over Mercury 2. What actually moved, why the launch price is a promotion, and how to test it.
Read
ModelsMiMo-V2.5-Pro: A 1T Open Model and What Its 1M Window Delivers
Xiaomi's April 2026 flagship: 1.02T total, 42B active, text-only, MIT weights. What its long-context results show, and why V2.6-Pro is the default for new work.
Read
ModelsMiMo-V2.5: Xiaomi's Omnimodal MoE at 15B Active Parameters
MiMo-V2.5 reads text, images, video and audio from 310B total and 15B active parameters, under MIT. How close it gets to V2.5-Pro, and when it is still the pick.
ReadModelsMiMo-V2.6-Flash: 15B Active Parameters, Close Behind Pro
Xiaomi's MiMo-V2.6-Flash tracks V2.6-Pro within a few points on most agentic benchmarks at a third of the price. Where the gap opens, and which checkpoint to pick.
Read
ModelsMiMo-V2.6-Pro-UltraSpeed: When 10x the Price for Speed Pays
UltraSpeed is the fast edition of MiMo-V2.6-Pro, listed at ten times the standard rate. What is known, the cost arithmetic, and how to measure the speed you buy.
ReadModelsMiMo-V2.6-Pro: What Scaled RL Changed in Xiaomi's Flagship
Same 1.02T/42B size class as V2.5-Pro, now omnimodal, MIT-licensed and far stronger on agentic benchmarks. How to read Xiaomi's numbers and what to test first.
ReadShowing 181–192 of 446 articles