OLMoE Instruct

weights

open(Apache-2.0)

data

open(OLMoE-mix-0924 pretraining + Tulu/UltraFeedback post-training)

code

open(OLMoE training code, SFT/DPO/KTO recipes)

checkpoints

open(244 intermediate)

license

Apache-2.0(OSI)

Fully-open MoE: Apache-2.0 weights + open pretraining data (OLMoE-mix-0924, on Dolma/DataComp) + open post-training data + open code + 244 intermediate checkpoints. Open_source tier 5.

• https://huggingface.co/allenai/OLMoE-1B-7B-0924-Instruct recorded 2026-06-25

  License Apache-2.0; SFT+DPO instruct; open data and code

• https://allenai.org/blog/olmoe-an-open-small-and-state-of-the-art-mixture-of-experts-model-c258432d0514 recorded 2026-06-25

  released with open data, code, evaluations, logs, and intermediate training checkpoints

~37.9k HF downloads/month on the instruct SKU; 96 likes. Solid reach for a fully-open research MoE.

• https://huggingface.co/allenai/OLMoE-1B-7B-0924-Instruct recorded 2026-06-25

  ~37,880 downloads last month; 96 likes

State-of-the-art among ~1B-active-parameter models at release (beats Gemma2, Llama2-13B-Chat, OLMo-7B, DeepSeekMoE-16B per the blog), but absolute capability is low-tier vs full-size 2026 chat models given the 1.3B active budget. Scored 2 on absolute capability. Exact per-competitor numbers live in arXiv 2409.02060 (blog charts are images), so the comparative claims are directional -- hence medium confidence.

• https://huggingface.co/allenai/OLMoE-1B-7B-0924-Instruct recorded 2026-06-25

  MMLU 51.9, GSM8k 45.5, HumanEval 54.8, AlpacaEval 84.0, IFEval 48.1

• https://allenai.org/blog/olmoe-an-open-small-and-state-of-the-art-mixture-of-experts-model-c258432d0514 recorded 2026-06-25

  positioned state-of-the-art at the ~1B-active cost class