Aiqre

Model library / deepgrove

maple-preview

Maple-Preview is a 20.2 billion-parameter ternary-weight reasoning model from DeepGrove, supporting 131k token context and released under MIT.

20.2B

Mixture of experts, count not published

131,072

Context length in tokens, as published

38 GiB

Weights at the released precision

mit

Licence declared on the repository

Overview

What this model is.

Maple-Preview is a 20.2 billion-parameter, ternary-weight reasoning language model released by DeepGrove in 2026. The publisher describes it as an efficient on-device inference model aimed at strong reasoning tasks, including IMO-level problem solving, and notes it runs at over 200 tokens per second on a Mac mini M4.

The checkpoint is 5.31 GB and stored in bfloat16 format. It uses a 24-layer, 256-expert architecture with eight active experts and a 131,072-token context window. The model is licensed under MIT, allowing unrestricted use, and requires a CUDA environment with Triton and FlashAttention for the provided Transformers implementation.

Specification

The published shape.

Repositorydeepgrove/maple-preview
Publisherdeepgrove
Published4 August 2026
ArchitectureMapleForCausalLM
Parameters20,214,030,336
Layers24
Hidden size2,048
Attention heads16, 4 key/value heads
Context length131,072 tokens
Vocabulary151,936 tokens
Weight formatbfloat16
Licencemit

Memory

How much VRAM the weights need.

Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.

PrecisionWeightsNotes
bf1637.7 GiBFull precision as released
fp818.8 GiB8-bit, near-lossless on most models
int49.4 GiB4-bit, smallest footprint

The cache costs 48 KiB per token at 16-bit, so the full 131,072-token context of one request needs about 6.0 GiB. Concurrency multiplies that number, not the weights.

That figure is an upper bound: this model uses sliding-window attention with a 512-token window on some layers, so long requests cache less than this.

Hardware

Which card runs it.

At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.

GPUCards neededFree for cacheContext it holds
RTX 4090 24GB 2 0.5 GiB about 11,980 tokens
RTX 5090 32GB 2 14.9 GiB the full 131,072 tokens
L40S 48GB 1 3.0 GiB about 66,594 tokens
A100 80GB 1 31.8 GiB the full 131,072 tokens
H100 80GB 1 31.8 GiB the full 131,072 tokens
RTX PRO 6000 96GB 1 46.2 GiB the full 131,072 tokens
H200 141GB 1 86.7 GiB the full 131,072 tokens

Serving

Running it yourself.

A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.

vllm serve deepgrove/maple-preview \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.90

Questions

The things people ask about this model.

How much GPU memory does maple-preview need?

The weights take about 38 GiB at the released precision. That fits on one L40S 48GB, which leaves roughly 3 GiB for the key-value cache.

What licence does maple-preview use?

The repository declares mit. Read the licence text before commercial use: the name alone does not tell you what is allowed.

How long a context does it support?

The configuration allows 131,072 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.

Can I use it through an API instead of hosting it?

This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.

Also in the library

Related models.

Gryphe/Pantheon-Reasoning-26B-A4B-1.1-V2

Pantheon-Reasoning-26B-A4B-1.1-V2 is a 26.5 billion-parameter Gemma-4 mixture-of-experts finetune aimed at role-play scenarios, i…

nex-agi/Nex-N2.5-mini

Nex-N2.5-mini is a 35.1 billion-parameter mixture-of-experts language model released under Apache-2.0.

inclusionAI/LLaDA2.2-mini

LLaDA2.2-mini is a 16.3 billion-parameter MoE diffusion language model with 128 k token context, supporting Levenshtein editing.

IFM/K2-Horizon-32B

K2-Horizon-32B-Stage1 is a 32-billion-parameter dense decoder-only model released by IFM for benchmarking agentic, coding and rea…

Hosting

Want this model on a dedicated EU GPU?

Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.