Yi model series: Yi and Yi-1.5 compared
Yi is 01.AI's bilingual English and Chinese open language model line. Compare Yi and Yi-1.5 by size, context length and license.
What Yi is and who makes it
Yi is a family of bilingual (English and Chinese) language models from 01.AI. The Yi model cards state that the models follow the Llama architecture pattern but are not derivatives of Llama: the datasets and training infrastructure were built independently. The Yi technical paper (submitted March 2024) describes 6B and 34B base models, chat versions, long-context variants of 200K tokens, depth-upscaled versions and vision-language models, with a pretraining corpus of 3.1 trillion tokens of English and Chinese text. The cards for the generations below list Apache 2.0, which permits personal, academic and commercial use.
The generations in order
Yi (November 2023)
Yi opened with Yi-6B and Yi-34B, released on November 2, 2023. The 34B card lists 3 trillion training tokens, a default context length of 4K tokens that can be extended to 32K at inference, and Apache 2.0. According to the GitHub repository, 200K-context versions followed on November 5, 2023, the Yi-VL vision-language models (34B and 6B) in January 2024, and Yi-9B on March 6, 2024. Yi-9B was built by continuing training of Yi-6B on 0.8 trillion more tokens, aimed at code and math. The repository also lists 4-bit AWQ and 8-bit GPTQ quantized builds for consumer GPUs.
Yi-1.5 (May 2024)
Yi-1.5 was released on May 13, 2024 as an upgraded version of the same architecture. The card describes continued pretraining of Yi on a further 500B tokens of high-quality data, followed by fine-tuning on 3M samples, and states stronger coding, math, reasoning and instruction following with language understanding retained. Sizes are 6B, 9B and 34B, in base and chat versions. Context variants are 4K, 16K and 32K: the 9B and 34B 32K base models and the 16K chat models are separate checkpoints on the Hugging Face organization page. License is Apache 2.0.
We found no newer generation in the Yi language model line in the sources we opened. We did not search exhaustively, so treat that as "not found", not "does not exist".
One detail from the paper that matters for fine-tuners: the chat models were finetuned on fewer than 10,000 verified instruction examples, and the 200K context came from continual pretraining of the base model, not from the original training run. The card for the 34B model gives a training data cutoff of June 2023, so the original Yi has no knowledge of anything after that point, and Yi-1.5's extra 500B tokens are the only later data we could source.
Which generation to use today
- Almost any new project: Yi-1.5 over Yi. It keeps the architecture, adds the extra pretraining, and the sizes (6B, 9B, 34B) cover a single small GPU up to a large single-GPU or multi-GPU host.
- Long documents: pick the context variant on purpose. Yi-1.5 tops out at 32K in the checkpoints we saw. The original Yi line has 200K versions, which is the longest context in the family, so use those if you need it and accept the older base.
- Smaller than 9B: Yi-1.5-6B, or Yi-6B if you specifically need its 200K variant.
- Chinese and English bilingual workloads: either generation, as both are trained on English and Chinese.
- Vision input: Yi-VL, which belongs to the first generation.
Because the models date from 2023 and 2024, check how a candidate compares with a current model in your size range on your own prompts before you commit. Any benchmark ranking in the Yi paper or cards is 01.AI's own report, and the paper credits data quality from its data-engineering work as the main driver of results.
Running it
To try one, use Aquanode GPUs and read the memory estimate on the generation hub first.
Yi generations
Every Yi generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Yi-1 | 5 | 6.1B to 34.4B | 13.5 GB | RTX A4000 | $0.167/hr |
| Yi-1.5 | 6 | 6.1B to 34.4B | 13.5 GB | RTX A4000 | $0.167/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Best models by task
Yi models appear on these ranked task pages.
Sources
- https://huggingface.co/01-ai/Yi-34B
- https://github.com/01-ai/Yi
- https://arxiv.org/abs/2403.04652
- https://huggingface.co/01-ai/Yi-1.5-34B-Chat
- https://huggingface.co/01-ai/Yi-1.5-9B-32K
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.