# Hugging Face — Apriel-H1: The Surprising Key to Distilling Efficient Reasoning Models

- Company: Hugging Face (huggingface.co)
- Announced: 2025-11-19T05:19:07+00:00
- Category: new-model
- Subject: Platform
- Models affected: Apriel-H1-15b-Thinker-SFT
- Source: https://huggingface.co/blog/ServiceNow-AI/apriel-h1
- Record: https://forck.live/items/1585-apriel-h1-the-surprising-key-to-distilling-efficient-reasoning-models

ServiceNow AI converted their 15B reasoning model to a Mamba hybrid (Apriel-H1-15b-Thinker-SFT) achieving 2.1x throughput with minimal quality loss, using reverse KL divergence and staged distillation on high-quality reasoning traces.

## Evidence

Verbatim from https://huggingface.co/blog/ServiceNow-AI/apriel-h1:

> We converted our 15B reasoning model to a Mamba hybrid achieving 2.1x throughput with minimal quality loss.

---

Record: https://forck.live/items/1585-apriel-h1-the-surprising-key-to-distilling-efficient-reasoning-models
Catalogue: https://forck.live/llms.txt
Feed: https://forck.live/feed.md