ARCHITECTURELM

Models

Tools

Skills

About

PRACTICE

DELIVERY

LANGUAGE

URBAN

RENDERING

COMPUTATION

Fig. 01 — Source: www.lesswrong.com

No. 001

EDITORIAL

Welcome to ArchitectureLM

A publication and reference for practicing architects, organized around the question of who answers for the output when a machine helped make it.

Latest News - Stay current on AI in architecture

release

October 15, 2026

Meta open-sources Muse Spark Open, an 816B model that runs on two Mac Studios

816B parameters with 42B active, a 1M context, natively multimodal, weights on Hugging Face and the major clouds from 15 October under Meta's own Muse Model License. The desktop demonstration pairs two M5 Ultra Mac Studios at 512GB each, so running it locally is a hardware budget question before it is a confidentiality one. · Meta

↗

release

October 15, 2026

Black Forest Labs open-sources a 5.6B image model that runs on 8GB of VRAM

5.6B parameters, editing and up to ten reference images, weights on Hugging Face under the FLUX Kommunity License, with an INT8 build that runs on 8GB of VRAM. Free commercially below $5m turnover, which puts the self-hosting decision on a practice's own revenue line, not a vendor's API terms. · Black Forest Labs

↗

update

October 14, 2026

OpenAI sets 14 January for the GPT-4o and GPT-4.1 API shutdown

gpt-4o-2024-08-06, gpt-4o-2024-11-20 and gpt-4.1-2025-04-14, plus the mini snapshots, off the API on 14 January 2027, with gpt-6-luna named as the replacement. Any plugin or script in the office that pins a model string has three months, and nobody owns that check by default. · OpenAI

↗

release

October 13, 2026

Liquid AI posts LFM2.5-24B-A2B weights under a licence capped at $10m revenue

24B parameters with 2B active, a 131K context, and GGUF checkpoints on Hugging Face under Liquid's own LFM Open License v1.0. Free commercially below $10m annual revenue — which is where a practice weighing self-hosting against a vendor API should read the licence. · Liquid AI

↗

release

October 13, 2026

Google ships Gemini 4 Argon at a rate that doubles after 30 November

$2 and $10 per million tokens, a 1M unified input-output limit, and cached input at 95% off — then $4 and $20 from December. Any fee priced against the introductory rate needs rechecking before the pilot becomes a workflow. · Google

↗

release

October 12, 2026

Meta ships Muse Wonder on the API and keeps Muse Wonder Full to subscriptions

Available 12 October at $4 and $20 per million tokens as muse-wonder-1.0, with a one-million-token context and no public weights; the stronger Muse Wonder Full is Muse Code subscribers only, at no published API rate. Anything costed on the headline rate is costed on the lesser model. · Meta

↗

release

October 9, 2026

Alibaba's Qwen posts Qwen-Image-2.1-Turbo weights under a research-only license

An accelerated 7B checkpoint of Qwen-Image-2.1 that generates and edits in eight denoising steps, weights on Hugging Face from 9 October, 2K output. Non-commercial use only, so a practice self-hosting it for client visuals needs a separate license from the lab. · Hugging Face

↗

new

October 8, 2026

Magnific ships Magnific One with an optional brand kit that checks final images

Live on all paid plans from 8 October, reachable through its MCP, with a Draft grid of 8 or 16 before a 2K or 4K final. Magnific's own page names no base model, so what actually renders the image is not on the product page. · VentureBeat

↗

All Articles

More is on the way.

Articles are being written and will be filed here as they are ready. In the meantime the reference indexes are open — models, tools and skills, each with a verdict attached rather than a vendor summary.

Models ↗

Tools ↗

Skills ↗

Newsletter

Five minutes a week on what actually changed.

PUBLICATION

ArchitectureLM

SUBJECT

AI in architecture

NEWSLETTER

Weekly, free

CONTACT

hello@architecturelm.com

Contact

Privacy

Acceptable Use

Terms