Yi-1.5

What Yi-1.5 is

Yi-1.5 is a family of open-weight large language models released by 01.AI on 13 May 2024. It comes in three parameter sizes — 34B, 9B and 6B — and each size is published in two forms, a base model and a chat model^[1][3]. The base models are the raw pre-trained networks, meant to be adapted or fine-tuned further; the chat models are the versions tuned to hold a conversation and follow instructions, and they are the ones most people would use directly.

Where it came from

Yi-1.5 is an upgrade of the earlier Yi models rather than a design built from scratch. 01.AI took Yi and kept training it — an approach usually called continued pre-training — on a high-quality corpus of 500 billion tokens, and then fine-tuned the result on 3 million diverse samples^[2]. Continued pre-training means the existing weights are not thrown away: the model is fed a large amount of additional text so that it absorbs more knowledge and language patterns, which is generally cheaper than training a model of the same size from zero. The overview table in the model card puts the total pre-training budget at 3.6 trillion tokens^[2].

Sizes, context windows and license

The family covers a wide range of deployment budgets. The 34B models suit settings where a single large GPU server is available, while the 9B and 6B models are small enough to be practical in more modest environments. Context length is treated as a separate choice: the model card lists context variants of 4K, 16K and 32K tokens^[2], and 01.AI also released longer-context versions alongside the standard ones, such as Yi-1.5-34B-32K^[2]. A longer context window lets a model take in more text — a long document or a long conversation — before it has to drop earlier material.

Code and weights are distributed under the Apache 2.0 license, which allows commercial use^[1]. For companies that want to build on a model without negotiating a separate agreement, that licensing choice matters as much as the benchmark numbers.

Reported results

01.AI's own comparisons focus on the small chat models. The company said its 6B and 9B chat models beat other models of the same size as well as Llama-3-8B-Instruct on GSM-8K, MATH, HumanEval and MBPP, and led same-size models on MT-Bench, AlignBench and AlpacaEval^[3]. The first group of benchmarks covers grade-school maths, competition-style maths and code generation; the second group covers general chat quality and alignment judged by models or preference data. 01.AI also said that Yi-1.5-9B-Chat was the top performer among similarly sized open-source models^[2]. These are the company's own evaluations, and independent testing can rank models differently depending on prompts and settings.

Place in the line

Yi-1.5 did not appear on its own. It was launched alongside the closed Yi-Large model and the Wanzhi app at 01.AI's first-anniversary event^[3][4], which framed the release as a two-track strategy: openly licensed small and mid-sized models for outside developers, and a large proprietary model reached through a product. The weights were collected on Hugging Face, where the Yi-1.5 collection was updated on 20 May 2024^[2], a week after the launch. Yi-1.5 therefore sits in the middle of the Yi series — a refresh of the original open models, released at the same moment the company moved into much larger closed models.