What Dominated the AI Landscape in Q2 2024? From Llama 3 to GPT‑4o and Global Price Wars

Meta releases open‑source Llama 3

On April 19, Meta announced Llama 3, its next‑generation open‑source large language model trained on a 24K‑GPU cluster with 15 TB of data. Two versions are offered: a 8 B‑parameter pre‑trained model and a 70 B‑parameter instruction‑tuned model, which Meta claims are the best in their size range. Improvements in post‑training reduced refusal rates, increased alignment, and boosted capabilities in reasoning, code generation, and instruction following.

Microsoft’s WizardLM‑2 appears then disappears

Just before Llama 3’s launch, Microsoft quietly released the open‑source WizardLM‑2 series (7 B, 70 B, and an 8×22 B flagship with 141 B parameters). Hours later the model was withdrawn because the team missed a required toxicity test. Microsoft later explained the omission as an oversight and promised to re‑release after proper testing.

Chumen Wenwen goes public

AI startup Chumen Wenwen listed on the Hong Kong Stock Exchange on April 24 at HK$2.98 per share, valuing the company at roughly HK$44.5 billion. The firm focuses on generative AI and voice interaction and is noted as one of the few Asian companies capable of building general‑purpose large models.

Google lays off its entire Python team

On April 27, Google announced that its Python engineering team, including core CPython contributors, was reassigned to other groups abroad rather than being directly terminated. The team had been responsible for maintaining the upstream Python codebase, the monorepo build system, and large‑scale refactoring tools.

Rabbit R1 hardware controversy

Tech blogger Mishaal Rahman reported on May 1 that the Rabbit R1 AI device runs a standard Android OS, with its “AI” functionality delivered via a regular Android app packaged as an APK, not a dedicated AI chipset.

Alibaba Cloud’s Qwen 2 and aggressive pricing

On June 7, Alibaba Cloud released the open‑source Qwen 2‑72B model, claiming performance surpasses Llama 3‑70B and other leading Chinese models. Earlier in May, Alibaba reduced Qwen Max to ¥0.04 per 1k tokens (‑67%) and Qwen‑Long to ¥0.0005 per 1k tokens (‑97%).

OpenAI unveils GPT‑4o

OpenAI introduced GPT‑4o on May 14, a multimodal flagship model that adds voice, image, and audio capabilities to the existing GPT‑4‑level intelligence. The model can understand speech, generate spoken responses with controllable tone, and perform visual reasoning.

Google’s Gemini 1.5 series and Project Astra

At the I/O conference on May 15, Google announced Gemini 1.5 Pro, Gemini 1.5 Flash, Gemini Advanced, and the AI agent Project Astra for smartphones, along with new media‑generation models Veo and Imagen 3. The updates aim to compete with OpenAI’s Sora and DALL‑E 3.

Price wars across Chinese AI providers

From May 21 to early June, Alibaba Cloud, Baidu Cloud, Tencent Cloud, and iFlytek announced steep price cuts or free tiers for their large‑model APIs, driving token costs down to fractions of a cent and prompting a “price‑war” narrative among users.

Mistral releases Codestral for code generation

On May 29, French AI startup Mistral, backed by Microsoft, launched Codestral, an open‑source generative model specialized in code completion and programming assistance, with high parameter counts and licensing restrictions.

Stanford‑led “Llama 3‑V” controversy

Stanford researchers claimed to have built a multimodal model, Llama 3‑V, that outperforms GPT‑4V while costing only $500 to train. Subsequent investigation revealed the model was a repackaged version of Tsinghua‑affiliated MiniCPM‑Llama3‑V 2.5, sparking debate over academic attribution and “model‑shelling.”

New AI ventures and policy shifts

Former OpenAI chief scientist Ilya Sutskever co‑founded Safe Superintelligence (SSI) on June 19, focusing on safe, high‑performance AI without immediate commercial intent.

The U.S. House passed the ENFORCE Act on June 23, tightening export controls on AI technologies and potentially restricting H‑1B holders from working on AI/ML projects.

Anthropic’s Claude 3.5 Sonnet and other model releases

Anthropic released Claude 3.5 Sonnet on June 20, claiming modest performance gains over GPT‑4o and its own Claude 3 Opus. Meanwhile, Huawei announced its domestically developed programming language “Cangjie” at the HDC 2024 conference, targeting cross‑platform development on HarmonyOS.

OpenAI service termination and Chinese “migration” plans

On June 25, OpenAI announced it would cease API services to unsupported countries—including China—effective July 9. Chinese providers such as Zhipu AI, Baichuan, and Tencent quickly launched “migration” programs offering massive free token allocations and discounted pricing to attract displaced developers.