Microsoft releases Phi-2, an AI language model the company says outperforms Mistral and Llama 2 at 7B and 13B parameters and is small enough to run on a phone
The competitive frame quickly broadened when Stability AI claimed its 1.6B model beat Phi-2 on most benchmarks, while Microsoft continued the family with Phi-4 mini and multimodal variants. The important arc is therefore sustained competition over capability per parameter.
First-order effects
Microsoft gains a compact-model benchmark claim against Mistral and Llama 2 systems in larger parameter classes, strengthening Phi-2's positioning for resource-constrained deployments.
Developers evaluating on-device language features have another Microsoft model explicitly positioned for phone-class execution, alongside its claimed performance advantage.
Second-order effects
Rivals in small language models face pressure to demonstrate not just low parameter counts but comparable task performance; Stability AI's subsequent claim for Stable LM 2 1.6B shows that benchmark contest emerging quickly.
For application builders, a credible smaller model expands the trade-off set between local inference and centrally hosted models, making responsiveness and privacy more relevant selection criteria where device execution is feasible.
Third-order effects
If compact models keep narrowing performance gaps with larger systems, model competition will increasingly center on efficiency, deployment fit, and distribution through devices—not parameter scale alone.
That shift could give platform owners that control operating systems and device integrations an advantage, while making benchmark claims less durable unless they translate into broadly usable on-device products.
The trend: AI model development is moving toward smaller, capable models designed to place inference closer to the user and inside device ecosystems.
Yesterday, new Mistral killed Gemini Pro. Today, Microsoft showed us that GPT4 Turbo with proper prompting is better than Gemini Ultra. Microsoft also launched Phi-2. A small LLM (2.7b) that beat Gemini Nano (3.2b), Mistral 7b, and beat llama 70b on coding.
Today, we share our teams' latest contributions, Phi-2 and promptbase. Phi-2 outperforms other existing small language models, yet it's small enough to run on a laptop or mobile device. https://www.microsoft.com/...
Phi-2 numbers, finally! We're seeing a consistent ranking: phi-2 outperforms Mistral 7B & Gemini Nano 2* (*on their reported benchmarks) and is roughly comparable to Llama 2-70B (sometimes better, sometimes worse). Beyond benchmarks, playing with the models tells a similar story.…
Microsoft just released Phi-2 (2.7B) which seems to come close to Llama 70B. It's trained on 1.4T tokens of heavily filtered web data and a lot of GPT 3.5/4 outputs. Knowledge distillation really does work, and combining it with other high quality data could be the key for OSS. […
Phi-2 was trained on 1.4T tokens, the training took 14 days on 96 A100 GPUs. They claim to have already surpassed Google Nano despite being a smaller size. [image]
Phi-2 A new large language model that beats its rivaling 25x larger counterpart. What else? It runs on your mobile device. Truly personal LLM's that work offline using edge compute is here https://www.microsoft.com/... [image]
More https://www.microsoft.com/... on promptbase & advances in prompting, including how we steered GPT-4 to record performance on the Measuring Massive Multitask Language Understanding (MMLU) benchmark, spanning 57 areas @satyanadella @kevin_scott @peteratmsr @HarshaNori @MSFTRes…
Contrary to rumors, I assure you the Phi-2 model is downloadable, but it is against the license to redistribute it. You can login to Azure ML Studio (free with a basic Azure account), select it from the Model Catalog, and then download the files from the Artifacts tab. [image]
From new best-in-class small language models to state-of-the-art prompting techniques, we're excited to share these innovations and put them in the hands of researchers and developers.
Microsoft just released Phi-2, a 2.7B LLM that rivals the 25x bigger LLaMa-2 70B. The best part? The model is small enough to run on a laptop or mobile device. Trained on 1.4T tokens: mixture of synthetic & web datasets, it beats Mistral 7B and Llama-2-70B model on muti-step... […