llm-meta-ai 0.2
Simon Willison’s Weblog 5 days ago 46 ● 27 sources
Meta AI released llm-meta-ai 0.2, adding support for its Muse Wonder and Muse Wonder Full models.
Large Language Models
442 summarised stories about Large Language Models, each linking back to the original source. Browse all topics →
Key players
Monday, 12 October 2026
Meta releases its first frontier model, Muse Wonder
TechCrunch 5 days ago 44 ● 27 sources
Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full as its first frontier model attempt. The Muse Wonder API pricing is $4.00 per million input tokens and $20.00 per million output tokens. The release adds a larger, higher-benchmark model for coding and powers the Muse agent, with Wonder Full limited to Muse Code subscribers and not available via the API while emphasizing “active memory” on MemoryBench.
Accessing OpenAI GPT models on Amazon Bedrock with global cross-Region inference in Thailand, Malaysia, Singapore, Indonesia, and Taiwan
Amazon Web Services 5 days ago 9
Amazon Bedrock added global cross-Region inference profiles so teams in Thailand, Malaysia, Singapore, Indonesia, and Taiwan can call OpenAI GPT-6 models even when local single-Region capacity is throttled. The profiles include GPT-6 Sol and GPT-6 Luna routed to ap-southeast-1 (Singapore) with cross-Region destinations such as ap-southeast-3 (Jakarta), ap-southeast-5 (Malaysia), ap-southeast-7 (Thailand), and ap-east-2 (Taipei). Calls are configured by using a global inference profile ID as the model, with permissions, quotas, and monitoring applied while billing and logging remain in the source Region.
Can we change an AI model’s architecture without retraining it from scratch?
Amazon Science 5 days ago 14
Amazon Web Services introduced “priming,” a method that reuses pretrained Transformer weights to convert parts of the architecture into hybrid state space/attention models without starting pretraining over.
Meta’s Muse Wonder Overtakes GPT-6 and Gemini 4
Trending Topics 5 days ago 28 ● 27 sources
Meta introduced Muse Wonder and Muse Wonder Full A.I. models that, in its benchmarks and an independent ranking, match or beat OpenAI’s GPT-6 and Google’s Gemini 4. Muse Wonder Full scored 55.8 points on Artificial Analysis’s Intelligence Index, about 3 points ahead of GPT-6 Astra and Gemini 4 Argon at the top compute setting. Meta’s top models remain closed with no released weights and are accessible only via Meta’s Model API and subscriptions, while only smaller models are open.
Deploy decision models with llama.cpp on Amazon SageMaker AI
Amazon Web Services 5 days ago 23
The post shows how to deploy decision models using llama.cpp on Amazon SageMaker AI, including an OpenAI-compatible text-generation endpoint and an optional SemIf extension for structured option probabilities. It benchmarks SemIf on 231 JevBench tasks and reports the GPU endpoint is 32.2 times faster at p95 than the CPU setup. As a result, you can route requests over a fixed set of options using probability outputs without generating an answer token, and choose CPU for seconds-level latency or GPU for tighter tail latency.
Why Isn't the Industry Freaking Out About DeepSeek 4.1 Flash?
dgt.is 5 days ago 22
The article describes the author’s month-long use of DeepSeek 4.1 Flash across many projects, arguing it feels as capable as top “frontier” models despite being much cheaper. It claims DeepSeek 4.1 Flash reduced KV-cache size by about 437x versus DeepSeek V1, which the author says keeps long coding sessions under $1 (and sometimes $0.003 per task). It changes developers’ behavior by making high-volume, low-stakes “unattended” work and long sessions economically practical, reducing reliance on expensive frontier subscriptions and suggesting future on-device/locally run efficiency improvements.
Build Your Own Decision Model
Another Dev's Two Cents 5 days ago 33
A developer showed how to build decision-model style multiple-choice inference with an LLM by constraining outputs to fixed option tokens and using their softmax scores as pseudo confidence. They tested Qwen/Qwen3-1.7B on a CommonsenseQA sample of 1221 items and reported 725/1221 accuracy (0.5938). The approach changed from overconfident, uncalibrated scores to better-calibrated probabilities after applying temperature scaling with a fitted temperature of 3.797280788421631, and they released scripts to reproduce evaluation, finetuning, and calibration.
Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged
Apple Machine Learning Research 6 days ago 21
The paper evaluated large language models as alternatives to human judgment for machine-translation quality annotation by comparing agreement with human annotators. It used a long-context test set covering 70 language pairs and also the WMT23 and WMT25 datasets. Results varied by evaluation scheme, language pair, and domain, so LLMs were not reliably consistent for most settings and the authors call for improvements, possibly via human–LLM collaborative annotation pipelines.