Model library / deepgrove
Maple-Preview is a 20.2 billion-parameter ternary-weight reasoning model from DeepGrove, supporting 131k token context and released under MIT.
Mixture of experts, count not published
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
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
| Repository | deepgrove/maple-preview |
| Publisher | deepgrove |
| Published | 4 August 2026 |
| Architecture | MapleForCausalLM |
| Parameters | 20,214,030,336 |
| Layers | 24 |
| Hidden size | 2,048 |
| Attention heads | 16, 4 key/value heads |
| Context length | 131,072 tokens |
| Vocabulary | 151,936 tokens |
| Weight format | bfloat16 |
| Licence | mit |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 37.7 GiB | Full precision as released |
| fp8 | 18.8 GiB | 8-bit, near-lossless on most models |
| int4 | 9.4 GiB | 4-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
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.
| GPU | Cards needed | Free for cache | Context 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
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 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.
The repository declares mit. Read the licence text before commercial use: the name alone does not tell you what is allowed.
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.
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 hello@aiqre.com with the model name and roughly what volume you expect.
Also in the library
Hy-Memory-30B-A3B is a Tencent-developed long-term-memory writer model designed to replace LLMs in Mem0, MemOS and Hy-Memory fram…
Granite Switch 4.1 30B is a modular LLM combining a 30-billion base, decision and guardian adapters, and an English speech-to-tex…
Sabi-Noema-MLA is a 27.4 B-parameter agentic model derived from Qwen3.8-27B, using multi-head latent attention to reduce attentio…
LFM2.5-24B-A2B is a 24 billion-parameter MoE reasoning model with 2.45 billion active parameters for on-device agentic tasks.
Hosting
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.