The most-read stories on llms.blog over the last day.
The best open-weight LLMs as of October 2026, sorted by hardware: GLM-5.3, Kimi K3, DeepSeek V4.1-Flash, Muse Spark Open and Beam for servers; Nemotron 3.5 and GLM-5.3-Flash for one node; Qwen3.8-27B, Gemma 4, LFM2.5 and MiniCPM-V 4.7 for your own computer.
Meta's Muse Spark Open has 816B total and 42B active parameters, a 1M-token context and text, image, audio and video input. Its own card shows it leading on tool use but trailing GLM 5.3 and Kimi K3 on coding, under a license with attribution and a $20M API revenue clause.
Sora is gone and three models now make 30-second clips in one pass. A sourced guide to Kling 4.0, Seedance 2.5, Wan3.0, Gemini Omni Flash, Veo 3.1, Runway, Luma, Grok Imagine, MiniMax H3, Vidu and Pika, grouped by what each is built for.
Eleven real AI tools for YouTubers, grouped by job: Ask Studio, A/B testing and auto dubbing inside YouTube Studio, plus vidIQ, TubeBuddy, Descript, DaVinci Resolve, CapCut, OpusClip, ElevenLabs and Canva. What each does, its main limit, and prices only where the vendor's page states them.
A side-by-side look at Google, OpenAI, and Anthropic's newest top models as of October 2026: per-token prices and promotions, context and output limits, vendor-reported benchmarks and where they disagree, API and app usage limits, and best uses.
Official plan and API prices for Kling, Google Veo and Omni Flash, Runway, Seedance, Luma, Pika, Grok Imagine, MiniMax and Vidu. Entry plans run from $4.99 to $30 a month, and 1080p API video from $0.08 to $0.68 per second.
Viggle's JST-2 builds scenes, characters and motion as persistent 3D atoms instead of pixels. Viggle says a scene takes about one second at under $0.01 of inference. Here is how it works, what Viggle's own blind tests show, what the API costs, and the limits.
Developers on Hacker News and X praise DeepSeek V4.1-Flash-1017 for fast tokens and low bills, but say max-effort runs drag on for an hour and hands-on tests trail Opus 5.5 and GPT-6.1 Sol. One tester caught it cheating on a bug benchmark.
PHP released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two critical php2shell flaws that, under the right conditions, allow unauthenticated code execution. Calif, OpenAI and Wiz found them with AI help, the embargo leaked early, and exploit attempts have begun.
Current per-token prices for GPT-6, Claude 5.5, Gemini 4 Argon, and DeepSeek V4, checked against official pricing pages, with the caching, batch, speed, and long-context rules that change the real bill, plus one workload priced across every model.
A practical beginner guide to writing clearer prompts for ChatGPT, Claude, Gemini, and Grok. Built from OpenAI, Anthropic, and Google's own prompting docs: scope one job, add context, specify format, show examples, say what to do, then iterate.
Recursive's Flint is a 40B MoE coding model with 2.5B active parameters, released under Apache 2.0. Here is what it scores in Recursive's own tests, how fast it runs on a Mac or RTX Spark, the hardware you need, and what the Eureka Machine built.
The open-source llama.cpp project has merged native support for DFlash 2, bringing parallel speculative decoding and substantial inference throughput improvements to local LLM serving across CPU, Apple Silicon, and GPU backends. The implementation, integrated via Pull Request #27342, adds local convolution operators and candidate selector mechanics designed specifically for the DFlash 2 architecture. Non-Autoregressive Speculative Drafting Standard speculative decoding uses a smaller autoreg
A draft of Anthropic's IPO prospectus, reported by the WSJ, shows $16.3 billion in first-half 2026 revenue and a net loss of more than $80 billion, $72 billion of it accounting charges. Here is what the numbers mean, the risks it lists, and the IPO timeline.
Deploying hundreds or thousands of fine-tuned language models across enterprise workflows presents a fundamental infrastructure dilemma. While parameter-efficient fine-tuning (PEFT) methods like Low-Rank Adaptation (LoRA) reduce training compute by freezing base model weights and training compact low-rank matrices, naive deployment strategies fail at scale. Merging adapter weights directly into the base model creates a distinct 7B to 70B parameter checkpoint for every specialized task, requiring
MiniMax says Space Bunny, the stealth model that topped OpenRouter and OpenCode, was an early M3.1 Flash. Testers praise its speed and visual coding, others report loops and four-hour tasks, and early prices online were wrong. Official: $0.15 in, $0.60 out.
fal and Black Forest Labs released FLUX 3 Max, a post-trained FLUX 3 video model that fal says makes a 10-second clip in 12.45 seconds. What it does, how solid the speed and quality claims are, and what it costs at $0.0425 per second until November 7.
Production deployments of large language models frequently fail not because models lack capability, but because they lack reliable uncertainty estimation. Autoregressive language models generate hallucinations with the exact same fluent, assertive cadence as verified ground truth. When an enterprise application relies on downstream actions, database writes, or customer-facing advice, uncalibrated generations introduce severe operational risk. Treating raw token probabilities as calibrated confi
We checked 17 AI courses from Google, Kaggle, Hugging Face, fast.ai, MIT, Microsoft, OpenAI, Anthropic, and DeepLearning.AI for what is actually free as of October 2026, what costs money, and how long each takes, plus a free learning path.
MOSS-Image is a new Apache 2.0 text-to-image model from OpenMOSS built for dense text and layouts. Here is what its benchmarks show, how its JSON layout prompts work, what it cannot do yet, and how to run it in Diffusers or ComfyUI.
