← All updates

Prometheus Supercluster Behind Muse

Plus: Anthropic Eyes $45B Compute Deal, Google Opens EmbeddingGemma 2

Good Morning, Model Builders!

MiniMax ships M3.1 Flash at a third of the cost, and Anthropic is in talks to buy $45 billion of compute from Nscale.


Meta Details Prometheus Supercluster Behind Muse

TL;DR: Meta says its Muse models, including Muse Wonder, were trained on Prometheus, its gigawatt-scale supercluster in Ohio with hundreds of thousands of GPUs. Meta reports over 85% goodput in pretraining, over 90% in reinforcement learning, and about 47,000 GPU hours between interruptions during the Muse Wonder run. Read more →


MiniMax M3.1 Flash Arrives

TL;DR: MiniMax released M3.1 Flash at one third the cost of M3 and three times the speed, scoring 74% on DeepSWE and 91% on BabyVision. API pricing is a flat $0.15 input and $0.60 output per million tokens, with cache reads at $0.015. Open weights are due on October 23. Read more →


Anthropic Eyes $45B Compute Deal

TL;DR: Anthropic is in talks to buy $45 billion of computing power from Nscale, the Nvidia-backed data center developer, according to Bloomberg. A deal could be finalized within days. It would be another large block of capacity for Claude training and inference as Anthropic keeps stacking compute across multiple suppliers. Read more →


Nvidia Launches Open Models for Agents

TL;DR: Nvidia launched new open models built to run AI agents inside enterprise data centers and on PCs. The push targets companies that want to host their own open models rather than call a closed API, and it extends Nvidia's bet on private, on-device agents that run locally on its hardware. Read more →


Google Opens Multimodal EmbeddingGemma 2

TL;DR: Google released EmbeddingGemma 2, a compact open-weight embedding model built on Gemma 4 under the Apache 2.0 license. It now embeds code, images, video and audio alongside text, aiming to make semantic search and retrieval-augmented generation cheaper to run for teams building their own RAG pipelines. Read more →


GoTo Releases 27B Indonesian Language Model

TL;DR: Indonesia's GoTo Group released a 27-billion-parameter language model trained for Indonesian and several regional languages, published through its official Hugging Face repository. It gives builders in Southeast Asia an openly downloadable model tuned for local languages that global frontier models often handle poorly. Read more →


Nvidia Open Sources Full Run:ai Platform

TL;DR: Nvidia released its entire Run:ai GPU orchestration platform as open source under Apache 2.0. The release includes the KAI Scheduler, CUDA 13.4 GPU fractioning, quota management and a local Kind dev environment, letting teams self-host GPU scheduling for model workloads on Kubernetes without a license. Read more →


Amazon Targets Nvidia With TorchNeuron

TL;DR: Amazon launched TorchNeuron, a new developer tool aimed at breaking Nvidia's hold on PyTorch, the toolkit used to build most AI models. The goal is to let more developers run their PyTorch workloads on alternative chips such as Amazon's own Trainium with less porting work. Read more →


Tenstorrent Raises $1.5B at $6.2B Valuation

TL;DR: Nvidia challenger Tenstorrent raised $1.5 billion at a $6.2 billion valuation after drawing acquisition interest earlier this year, choosing to stay independent. The money funds scaling its AI chips, adding capital to one of the few alternative accelerator suppliers outside Nvidia and AMD. Read more →


Report Says DeepSeek, Huawei Team on Chips

TL;DR: DeepSeek and Huawei are collaborating on new AI chip designs to challenge Nvidia, according to a report from The Information. Neither company has confirmed it. A tighter model and chip partnership would deepen China's push to train and serve frontier open models on domestic silicon. Read more →


Nvidia Opens Vera Rubin Rack Designs

TL;DR: At the 2026 OCP Global Summit, Nvidia opened its Vera Rubin rack designs, released open-source AI factory software, and announced plans with Cisco for a new pluggable optics specification. Open rack designs let more builders deploy Nvidia's next platform for model training and inference at scale. Read more →


Top AI models. Lower costs.

Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.