Maple-Preview: Native Ternary-Weight MoE for High-Speed On-Device Reasoning

DeepGrove has introduced Maple-Preview, an open-source 20B-A1B ternary-weight reasoning Large Language Model (LLM). The model is designed specifically for high-performance on-device inference, achieving speeds of 127 tokens per second on an iPhone and 218 tokens per second on a Mac mini M4, significantly outpacing other efficient models like Gemma 4 and Qwen3.5.

Native Ternary Weights for Inference Efficiency

Maple-Preview is trained natively in a ternary-weight format rather than being converted from a full-precision model. DeepGrove posits that converting pre-trained models to lower bitwidths unnecessarily limits both performance and efficiency. By treating low precision as a first-class citizen during training, the model can leverage the mathematical advantages of ultra-low bitwidths.

At ultra-low bitwidths, matrix multiplication can be effectively replaced with additions, which lowers the total arithmetic workload required for inference. This architectural choice allows Maple-Preview to maintain a small memory footprint (5.31 GB checkpoint) while supporting a 131,072-token context window.

Hardware-Aware Architecture

To optimize for on-device speed, the Maple architecture was developed through a hardware-aware design loop, with configurations tested directly on Mac mini hardware. The final architecture consists of:

Layer and Expert Configuration: A 24-layer, 256-expert configuration, chosen as a balance between reasoning performance and inference efficiency.
Attention Mechanism: A hybrid of sliding-window and global attention to limit the growth of the KV-cache.
Model Scale: 20.2B total parameters with 1.49B active parameters.

Reasoning Performance and Benchmarks

Maple-Preview is positioned as a reasoning-focused model, demonstrating strong capabilities in mathematical and technical domains. It successfully solved IMO 2024 Problem 1 (7/7) at 281.5 tokens/s on a MacBook Pro (M5 Pro).

In benchmark evaluations, Maple-Preview (Average score: 78.7) and its Flash variant (Average score: 77.7) compete effectively with larger models. For example, it outperforms GPT-OSS 20B (Average: 76.3) and Qwen3 30B-A3B (Average: 76.6) across a suite of tests including LCBv6, AIME 26, HMMT 26, and GPQA-D.

On-Device Adaptation via "Dreaming"

DeepGrove is experimenting with on-device weight adaptation to move beyond context-based memory. Instead of storing user preferences in text files or long contexts, the model can adapt its weights to retain subtle, user-specific details.

In a provided demo, the model identifies a user's vegan dietary restriction and schedules a "dream" session—a process where the model generates a small set of data and trains itself. This process takes 10–20 minutes with a peak memory utilization of 5.9 GB. Following this adaptation, the model can generalize the preference (e.g., recommending synthetic leather bags instead of real leather), a task where Claude Sonnet 5 with memory enabled reportedly failed.

Community Insights and Critiques

While the technical achievements in speed are widely praised, community members on Hacker News have raised several points regarding the model's reliability:

Hallucinations: Users reported that the model "hallucinates knowledge quite aggressively," particularly in niche areas outside of scientific or technical fields.
Knowledge Gaps: One user noted that the model was "confidently very incorrect" when asked about the etymology of specific words, suggesting that the 20B parameter size may limit its general knowledge base.
Tool Integration: Some developers suggest that the model's primary value lies in high-quality tool routing or as a fast backup model for smaller task sets, given its speed and limited internal knowledge.
Implementation Skepticism: Some users expressed skepticism regarding the "dreaming" mechanism, questioning whether a model of this size can effectively self-improve or if the process is limited to updating specific facts.