Model reference · open weights
maple is an open-weight language model from deepgrove. maple-preview (BF16) weighs 40.4 GB; the smallest configuration that runs it is L40S 48 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the deepgrove/maple-preview model card, 2026-10-01
What it is
| Released by | deepgrove |
|---|---|
| Released | 2026-08-04 |
| Parameters | 20.2B |
| VRAM | 40.4 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB … RTX 5090 32 GB 5 smaller cards | ||||
| L40S 48 GB | 9 | 4 | all 128K | 44.0 GB |
| A100 80 GB | 135 | 64 | all 128K | 78.2 GB |
| H100 80 GB | 77 | 30 | all 128K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 116 | 45 | all 128K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 150 | 58 | all 128K | 107 GB |
| H200 141 GB | 226 | 88 | all 128K | 138 GB |
| B200 180 GB | 316 | 79 | all 128K | 176 GB |
| 2× RTX 4090 24 GB tensor parallel | 15 | 7 | all 128K | 23.4 GB a card |
| 2× RTX 3090 24 GB tensor parallel | 15 | 7 | all 128K | 23.4 GB a card |
| 2× RTX 5090 32 GB tensor parallel | 71 | 33 | all 128K | 31.0 GB a card |
| 2× L40S 48 GB tensor parallel | 167 | 79 | all 128K | 44.0 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 41.8 GB | 42.1 GB |
| 5 | 42.9 GB | 44.4 GB |
| 8 | 43.7 GB | 46.1 GB |
| 16 | 45.8 GB | 50.7 GB |
| 32 | 50.2 GB | 59.8 GB |
| 64 | 58.8 GB | 78.2 GB |
One card, with vLLM's small-card settings.
From the model card
DeepGrove · 2026
Today we introduce Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM. Maple-Preview has SOTA reasoning for its weight class and is even competitive with larger models. It solves IMO-level problems and runs at 200+ tokens/sec on a Mac mini M4, 5–16× faster than efficient models like Gemma 4, Qwen3.5, and gpt-oss.
[!NOTE] The included Transformers implementation depends on Triton and FlashAttention and is intended for a compatible CUDA environment. The reported Apple Silicon result uses a separate on-device runtime.
Maple-Preview is a 20B-A1B reasoning model designed from the start for efficient on-device inference. It utilizes a 24-layer, 256-expert (8 active) configuration with 3:1 SWA-512:GA attention.
On benchmarks, Maple-Preview sets a new point on the Pareto frontier for both memory-to-performance and speed-to-performance, demonstrating its strong reasoning capabilities. However, we note that this preview is focused primarily on raw reasoning and, as such, may underperform on agentic benchmarks. We intend to continue improving general performance through extended training before Maple's full release.
Capability comparison using the dense output head across LCBv6, AIME 2026, HMMT 2026, and GPQA-D.
This preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning.
Maple-Preview is released under the MIT License.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.