China turns up the heat with open model blitz as US model makers panic

7 hours ago

Alibaba launches Qwen 3.8-Max, a 2.4 trillion-parameter open model, challenging US AI leadership.
DeepSeek releases V4 Flash 0731, a 284-billion parameter model that outperforms larger models with lower cost.
Chinese AI developers are competing on both price and performance, while US companies raise safety concerns.
Hugging Face CEO notes China's dominance in open models, likely extending to frontier AI soon.
Alibaba's model weights are now available for download, with a smaller 27-billion parameter version also planned.
DeepSeek V4 Flash offers significant cost savings, being 40% cheaper per task than OpenAI's GPT-5.6 Luna.
DeepSeek integrates DSpark speculative decoding to enhance efficiency, achieving faster inference without quality loss.
The model can run on high-end home hardware, like the DGX Spark, making powerful AI more accessible.

Moonshot AI released Kimi K3, a 2.8T parameter MoE model, with open weights scheduled for release on July 27th.
Kimi K3 is the strongest open model ever, ranking high on multiple benchmarks and competing closely with top closed models like Claude Fable and GPT-5.6 Sol.
The release reduces the performance gap between open/closed and U.S./Chinese models to an estimated 3-5 months.

<antirez>8 days ago

DwarfStar 4 (DS4) became popular quickly due to the release of a quasi-frontier model that is large and fast enough for local inference, combined with an asymmetric 2/8 bit quant setup that works on 96-128GB RAM.
The project leverages the experience of the local AI movement and GPT 5.5, allowing rapid development (one week) with proper LLM interaction skills.
The author worked 14 hours/day on average during the last week, similar to the early Redis days, but normal average is 4-6 hours.

OpenAI's GPT-Live enables simultaneous talking, listening, and formulating answers, improving ChatGPT banter.
SpaceXAI, formerly known for controversial AI, rebrands to offer legal and Excel automation services.
SambaNova's AI chip, backed by Intel, rejuvenates Nvidia GPUs, achieving 763 tok/s in benchmarks.

The user compares AI model pricing, noting that DeepSeek V4 is much cheaper than Anthropic and OpenAI's 'frontier' models, with a nearly 50x price difference based on tokens.
There's concern that Anthropic and OpenAI have high costs and may not be able to reduce prices significantly to compete with cheaper models like DeepSeek or Xiaomi's Mimo.
The user questions if lower costs for some models are due to being open-weight and stress-tested by many users, or if they're offered as loss leaders to drive prices down.

Moonshot AI released Kimi K3, a 2.8T parameter MoE model, with open weights scheduled for release on July 27th.
Kimi K3 is the strongest open model ever, ranking high on multiple benchmarks and competing closely with top closed models like Claude Fable and GPT-5.6 Sol.
The release reduces the performance gap between open/closed and U.S./Chinese models to an estimated 3-5 months.

<antirez>8 days ago

DwarfStar 4 (DS4) became popular quickly due to the release of a quasi-frontier model that is large and fast enough for local inference, combined with an asymmetric 2/8 bit quant setup that works on 96-128GB RAM.
The project leverages the experience of the local AI movement and GPT 5.5, allowing rapid development (one week) with proper LLM interaction skills.
The author worked 14 hours/day on average during the last week, similar to the early Redis days, but normal average is 4-6 hours.

OpenAI's GPT-Live enables simultaneous talking, listening, and formulating answers, improving ChatGPT banter.
SpaceXAI, formerly known for controversial AI, rebrands to offer legal and Excel automation services.
SambaNova's AI chip, backed by Intel, rejuvenates Nvidia GPUs, achieving 763 tok/s in benchmarks.

The user compares AI model pricing, noting that DeepSeek V4 is much cheaper than Anthropic and OpenAI's 'frontier' models, with a nearly 50x price difference based on tokens.
There's concern that Anthropic and OpenAI have high costs and may not be able to reduce prices significantly to compete with cheaper models like DeepSeek or Xiaomi's Mimo.
The user questions if lower costs for some models are due to being open-weight and stress-tested by many users, or if they're offered as loss leaders to drive prices down.

Hasty Briefsbeta