AI Potluck
Model components / Base / pretrained models

Yi-1.5

01.AI

01.AI's foundation model family at 6B, 9B and 34B with 32K-context variants, open sourced in May 2024. It continues pretraining from the original Yi release with a further 500 billion tokens, for 3.6 trillion in total.

Scores the base SKUs; the Chat releases are separate finetuned_chat records. A 2024-generation family, with adoption barely above its band floor. The benchmark figures this record once carried are no longer on the README or the model cards. Verified 2026-08-13 via the Yi-1.5-34B and Yi-1.5-9B model cards and the repository.

Openness

3 high confidence
3.0
weights
open(Apache-2.0, on HF)
data
closed(pretraining corpus not released)
code
partial(inference/usage examples
license
Apache-2.0(OSI, permissive, no use restrictions)

Permissive Apache-2.0 weights, but the training data and the full training code are not released, so this is open weights rather than open source.

Adoption

2 medium confidence
2.0

10,283 downloads in the trailing 30 days for the single declared artifact 01-ai/Yi-1.5-34B, which bands at level 2 (10K-100K) on the model adoption scale. It sits barely over the 10K floor - a 2024-era family superseded by the Qwen, Llama and DeepSeek generations - and one more month of decline would drop it a band.

Capability

2 medium confidence
2.0

Strong for a 2024 sub-35B model but well below the 2026 frontier open models such as DeepSeek-V4 and Qwen 3.6, which puts it alongside Falcon 3 among dated-generation families. The MMLU 76.3, BBH 54.3 and C-Eval 81.4 figures come from Table 2 of the Yi tech report and describe Yi-34B, the original 2023 model, rather than Yi-1.5-34B. No text source publishes those benchmarks for Yi-1.5-34B: the Hugging Face card carries only prose and chart images, and the 01-ai/Yi-1.5 README defers to the card. A 34B-Chat GSM8K figure of 90.2 that circulates elsewhere appears on none of those pages, so it is not carried here. The score of 2 holds on either release, since a family is scored on its strongest one and Yi-1.5 is the stronger.

  • https://arxiv.org/html/2403.04652v1 recorded 2026-08-14

    Yi tech report full text. Table 2, "Overall performance on grouped academic benchmarks compared to open-source base models", Yi-34B row - MMLU 76.3, BBH 54.3, C-Eval 81.4. Yi-6B row 63.2 / 42.8 / 72.0. These are Yi-34B figures, not Yi-1.5-34B.

  • https://arxiv.org/abs/2403.04652 recorded 2026-08-13

    Yi foundation-models tech report. The abstract states "Our base models achieve strong performance on a wide range of benchmarks like MMLU" but carries none of the recorded figures.

  • https://github.com/01-ai/Yi-1.5 recorded 2026-08-13

    Cited for Yi-1.5 base benchmark results (MMLU 76.3, BBH 54.3, C-Eval 81.4), but the README no longer carries a benchmark table - it says "For model details and benchmarks, see Model Card", and none of the three figures appears on either base card. The recorded value has no live source here.

Verified 2026-08-13