Good Morning, Model Builders!
DeepSeek opens up its inference engine, Meta brings capable agents to a single GPU, and new voice models push real-time AI forward.
DeepSeek Open-Sources Its Inference Engine

TL;DR: DeepSeek released DeepInfer, the vLLM-based inference framework that serves its fast V4.1-Flash model in production, as fully open source. The stack cuts compute cost and energy per token, which means anyone serving open models can copy DeepSeek's playbook and adds fresh pricing pressure on OpenAI and Anthropic. Read more →
US Orders DeepSeek Removal

TL;DR: The US government ordered federal civilian agencies to remove DeepSeek products, including third-party services built on its technology, citing national security risks. Vendors selling to Washington now need to check whether any DeepSeek model sits inside their stack, even when it arrives through an intermediary rather than DeepSeek's own app. Read more →
Ornith-1.5-122B Ships MIT Open Weights

TL;DR: Ornith released a 122B-parameter model under the MIT license that it says runs on a single NVIDIA DGX card, with SWE-bench Verified at 0.845 and top scores among comparably sized open models. Weights that capable, on one card, reset the floor for self-hosted coding and agent work. Read more →
DeepSeek Serves 155T Tokens a Day

TL;DR: DeepSeek's newly open inference stack runs a single cluster at 155 trillion tokens per day and $38 of revenue per GPU-hour on off-peak pricing, according to an analysis of the numbers DeepSeek published. Even at $7 per B300 GPU-hour that implies roughly 82 percent gross margin. Read more →
Meta's Muse Glimmer Runs on One GPU

TL;DR: Meta published a technical report for Muse Glimmer, a 30B open-weight model built for agentic tasks on consumer hardware. It runs at practical speed on a Mac or PC with one consumer GPU at 4-bit precision and leads comparably sized open models on tool use and multi-step planning. Read more →
Inception Ships Mercury 3 Diffusion LLM

TL;DR: Inception released Mercury 3, a diffusion language model it says generates 1,068 tokens per second with a 1M-token context window. On Inception's own table it beats GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks. Independent numbers are still pending, but the speed claim targets latency-bound agent workloads. Read more →
Mistral Launches Voxtral 3 Voice Stack

TL;DR: Mistral AI launched Voxtral 3, a model family for real-time voice agents that bundles speech-to-text, text-to-speech and a voice agent SDK. It supports more than 50 languages, can call external tools mid-conversation and can run on your own infrastructure, giving builders a self-hostable alternative to closed voice APIs. Read more →
ElevenLabs Unveils Faster Voice AI

TL;DR: ElevenLabs released Eleven Realtime, two speech-to-speech models built for live AI conversation. The v1 Max variant scores 98.8 percent on Big Bench Audio with a 0.57-second response time, and ElevenLabs says no model on the Artificial Analysis leaderboard is both faster and more accurate. 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.