1,107 Tokens Per Second: The LLM That Doesn't Type

·1 views

On September 8, 2026, Inception Labs announced Mercury 2.5 — which the company describes as the largest diffusion language model ever trained. The headline number: 1,107 tokens per second on widely available NVIDIA GPUs, at quality the company says matches the cost-optimized frontier tier (GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite, Claude Haiku 4.5). It's a 40% intelligence jump over Mercury 2, and it's aimed at one thing the whole industry pretends doesn't exist: the speed ceiling built into how every mainstream LLM writes. Every LLM you've ever used writes like a typewriter. It picks token 1,

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

[query-inspector] A Claude Code skill for extracting and tuning SQL/ORM queries

query-inspector is a Claude Code skill that extracts the SQL/ORM queries from your project's source, then diagnoses and tunes missing indexes / N+1 problems / anti-patterns. Queries that run fine until the data piles up. N+1 problems that sail through code review. The real SQL your ORM generates, invisible in the code. It catches all of this right before you commit, not after it reaches production.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why I Stopped Self-Hosting AI Models (And You Probably Should Too)

I spent three months and roughly $500 on GPU hardware trying to prove a point. That point was that I could run my own AI models, free from the shackles of API pricing and vendor lock-in. I was wrong, and the journey was both humbling and expensive. This isn't just a "cloud good, self-host bad" tirade. There are legitimate reasons to self-host—privacy, data sovereignty, or just the sheer nerdery of building your own inference rig.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Wirelist: cards, divided lists and grids for Laravel and Livewire

A list of cases, tasks or properties is not a table, as each record reads as a card, and forcing it into columns leaves gaps and odd widths. Wirelist draws those lists. Records as divided rows, separate cards or a grid, each one opening its record, with a row actions menu, load more and the 11 themes of the wire* family. It is built on Wiretables, so tables and lists share their look and sit together on one screen. It requires Wiretables, which Composer installs with it.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Part 6: Building it with MuleSoft Vibes: skills, rules, workflows and hooks

Part 6 of 10 · Building an Agentic Change-Approval MVP on MuleSoft Parts 4 and 5 covered what the integration team builds: MCP tool flows with strict contracts, behind a governed gateway. This part covers how they build it. We used MuleSoft Vibes in Anypoint Code Builder, and the interesting part wasn't the code it wrote. It was how we taught it our standards so every tool came out the same way. Vibes gives you three building blocks, and each one solves a different problem: What it is When it applies What we used it for Rules Natural-language constraints Every message (global or workspace) Sta

0 comments Read more at DEV Community