DeepSeek-V3 now runs at 20 tokens per second on Mac Studio, and that’s a nightmare for OpenAI

DeepSeek's free 685B-parameter AI model runs at 20 tokens/second on Apple's Mac Studio, outperforming Claude Sonnet while using just 200 watts, challenging OpenAI's cloud-dependent business model.Read More
Read more at VentureBeat
Topics
-
Mistral AI drops new open-source model that outperforms GPT-4o Mini with fraction of parameters
France's Mistral AI launches efficient open-source model that outperforms Google and OpenAI offerings with just 24 billion parameters, challenging U.S. tech giants' dominance in artificial ...VentureBeat - Mar. 17 -
Cerebras just announced 6 new AI datacenters that process 40M tokens per second — and it could be bad news for Nvidia
Cerebras Systems is challenging Nvidia with six new AI data centers across North America, promising 10x faster inference speeds and 7x cost reduction for companies using advanced AI models like ...VentureBeat - Mar. 11 -
Apple Mac Studio (M4 Max, 2025) Review: Small but Mighty
For creatives who need the very highest level of performance, the Mac Studio delivers it—and then some.Wired - Mar. 11 -
WATCH: Australian teen runs 200-meter dash in under 20 seconds
Australian teenager Gout Gout blew away the competition in the 200 meters, setting a wind-assisted time of 19.98 seconds, and set the fastest time in the world this year in the same heats.ABC News - Mar. 17 -
In AI Agent Battle, Meta Seeks Total Dominance as OpenAI Plans to Charge $20K for Some Models
If your company isn’t already using AI, the arrival of “agent” tools may change your mind. Meta wants this to happen, and OpenAI is planning on hefty prices for some advanced systems.Inc. - Mar. 7 -
Alibaba’s new open source model QwQ-32B matches DeepSeek-R1 with way smaller compute requirements
While DeepSeek-R1 operates with 671 billion parameters, QwQ-32B achieves comparable performance with a much smaller footprint.VentureBeat - Mar. 5 -
OpenAI urges U.S. to allow AI models to train on copyrighted material
OpenAI is asking the U.S. government to make it easier for AI companies to learn from copyrighted material, citing a need to “strengthen America’s lead” globally in advancing the technology.NBC News - Mar. 13 -
SimilarWeb data: This obscure AI startup grew 8,658% while OpenAI crawled at 9%
SimilarWeb data reveals dramatic AI market upheaval with Deepseek (8,658% growth) and Lovable (928% growth) dominating.VentureBeat - Mar. 5 -
What you need to know about Manus, the new AI agentic system from China hailed as a second ‘DeepSeek moment’
Manus AI is designed as a multi-agent system, meaning it combines several AI models to handle tasks independently.VentureBeat - Mar. 10
More from VentureBeat
-
Microsoft infuses enterprise agents with deep reasoning, unveils data Analyst agent that outsmarts competitors
Microsoft announced Tuesday two significant additions to its Copilot Studio platform: deep reasoning capabilities that enable agents to tackle complex problems through careful, methodical thinking, ...VentureBeat - 8h -
Beyond transformers: Nvidia’s MambaVision aims to unlock faster, cheaper enterprise computer vision
Nvidia is updating its computer vision models with new versions of MambaVision that combine the best of Mamba and transformers to improve efficiency.VentureBeat - 12h -
Gunzilla Games acquires, resurrects Game Informer
Game Informer returns as Gunzilla Games has acquired and relaunched Game Informer, bringing back the staff and the website.VentureBeat - 12h -
METASCALE improves LLM reasoning with adaptive strategies
METASCALE uses a three-stage approach to dynamically choose the right reasoning technique for each promblem.VentureBeat - 13h -
Google releases ‘most intelligent model to date,’ Gemini 2.5 Pro
Gemini 2.5 Pro is now available for Gemini Advanced users and is Google's most capable model with a 1 million token context window.VentureBeat - 14h
More in Tech
-
I Went Undercover in Crypto’s Answer to ‘Squid Game.’ It Nearly Broke Me
I spent 10 days competing in Crypto: The Game, a winner-takes-all contest where hundreds of players try to finesse and backstab their way to claiming a $140,000 cryptocurrency prize.Wired - 46m -
I Went Undercover in Crypto’s Answer to ‘Squid Game.’ It Nearly Broke Me
I spent 10 days competing in Crypto: The Game, a winner-takes-all contest where hundreds of players try to finesse and backstab their way to claiming a $140,000 cryptocurrency prize.Wired - 46m -
The Worm That No Computer Scientist Can Crack
One of the simplest, most over-studied organisms in the world is the C. elegans nematode. For 13 years, a project called OpenWorm has tried—and utterly failed—to simulate it.Wired - 1h -
The Best Programming Language for the End of the World
Once the grid goes down, an old programming language called Forth—and a new operating system called Collapse OS—may be our only salvation.Wired - 1h -
101 Best Amazon Spring Sale Deals (2025)
Now’s your chance to save on our favorite WIRED-tested home and tech gadgets.Wired - 2h