Model Details
Maple-Preview
Quick answer: Maple-Preview is DeepGrove's open-source (MIT-licensed) 20B-A1B reasoning model, released August 4, 2026, built as a natively-trained ternary-weight (2-bit, {-α, 0, +α}) network rather than a post-hoc quantized model. It scores 87.5% on AIME 2026, 78.8% on HMMT 2026, 75.1% on LiveCodeBench v6, and 73.5% on GPQA-Diamond (78.7% average across these four), while running at 218 tokens/s on an Apple Mac mini M4 — 5–16x faster than comparably-capable dense/MoE models like Gemma 4, Qwen3.5, and gpt-oss. The full checkpoint is 5.31 GB, with a 131,072-token context window.
At a Glance
Where Maple-Preview leads
- 218 tok/s decode speed on a Mac mini M4 — 5–16x faster than comparison models at similar capability, the core differentiator of its natively-ternary architecture
- 87.5% AIME 2026 and 78.8% HMMT 2026 — strong competition-math reasoning, on par with or ahead of several larger dense/MoE models
- 5.31 GB checkpoint — small enough to run entirely on-device (Mac mini, MacBook, even iPhone) at usable speeds
- Solves IMO-2024-level olympiad problems, demonstrated live at 281.5 tok/s on a MacBook Pro (M5 Pro)
- Fully open-source (MIT) — weights, conversion tooling, and inference code all public
Where it lags
- 82.9% average for Qwen3.5 35B-A3B (a larger model) still edges out Maple-Preview's 78.7% average, particularly on GPQA-Diamond (84.2% vs 73.5%)
- Minimal agentic post-training in this preview release — DeepGrove explicitly warns it may underperform on agentic benchmarks
- Only small-scale general reinforcement learning so far; full release is expected to improve further
Best for: on-device reasoning and math/coding tasks where raw decode speed and a small memory footprint matter — not yet recommended for agentic workloads pending DeepGrove's full release.
What Maple-Preview Is
Maple-Preview is DeepGrove's first public model, built to test a specific bet: that ultra-low-precision (ternary, 2-bit) networks should be trained natively at that precision from the start, rather than trained in full precision and quantized afterward. The architecture is a 20B-A1B (20.2B total / 1.49B active) mixture-of-experts model with 24 layers and 256 experts (8 active per token), using a hybrid 3:1 sliding-window (512-token) to global-attention ratio to bound KV-cache growth at long context lengths. Weights are stored as 2-bit packed ternary values ({-α, 0, +α}, one scale factor α per row), which lets matrix multiplication be replaced by additions during inference — substantially lowering the arithmetic workload compared to standard quantized models.
The practical payoff is decode speed: Maple-Preview reaches 218 tokens/s on a consumer Apple Mac mini M4 (using a "flash head" optimization that clusters vocabulary and only exactly scores the most likely clusters), and the DeepGrove team demonstrated it solving IMO 2024 Problem 1 correctly (7/7) at 281.5 tokens/s on a MacBook Pro with an M5 Pro chip. In a head-to-head demo, Maple-Preview generated a full recipe response on an iPhone in about 10 seconds (127 tok/s) versus nearly six minutes for a 1-bit-quantized 27B Qwen3.6-based comparison model.
This is explicitly a preview release focused on raw reasoning capability (math and coding), not agentic tool use — DeepGrove notes minimal agentic post-training and only small-scale general RL so far, with plans to scale agentic training and reinforcement learning ahead of a full release. The company also demoed an early on-device weight-adaptation ("dreaming") feature, where the model fine-tunes itself on a small locally-generated dataset to internalize user preferences — an alternative to growing context-based memory that DeepGrove argues generalizes better to subtle, adversarial follow-up requests.
Specifications
| Field | Value |
|---|---|
| Organization | DeepGrove |
| Parameters | 20.2B total / 1.49B active (20B-A1B) |
| Architecture | Ternary-weight (2-bit) MoE — 24 layers, 256 experts (8 active), 3:1 sliding-window (512 tok) : global attention |
| License | MIT |
| Release date | August 4, 2026 |
| Modality | Text |
| Context window | 131,072 tokens |
| Checkpoint size | 5.31 GB (2-bit packed) |
| Decode speed | 218 tok/s (Mac mini M4), 281.5 tok/s (MacBook Pro M5 Pro), 127 tok/s (iPhone) |
Pricing
Open weights under the MIT license — no per-token API pricing. Available on Hugging Face (full BF16 checkpoint) and as a pre-converted 2-bit MLX build (deepgrove/maple-preview-2bit-mlx) for Apple Silicon via deepgrove-ai/mlx-lm-deepgrove. The Transformers implementation requires Triton and FlashAttention in a CUDA environment; the Apple Silicon path uses a separate MLX-based runtime.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| LiveCodeBench v6 | 75.1% | DeepGrove launch page | 2026-08 |
| AIME 2026 | 87.5% | DeepGrove launch page | 2026-08 |
| HMMT 2026 | 78.8% | DeepGrove launch page | 2026-08 |
| GPQA Diamond | 73.5% | DeepGrove launch page | 2026-08 |
A separate "Flash" output-head variant (using the flash-head vocabulary-clustering optimization for extra decode speed) scores 75.6% LCBv6, 85.8% AIME26, 80.3% HMMT26, and 69.2% GPQA-D — DeepGrove's own published comparison uses the standard (non-Flash) head as the primary reported number.
Maple-Preview vs Alternatives
| Model | Total / Active Params | LCBv6 | AIME26 | HMMT26 | GPQA-D | Average |
|---|---|---|---|---|---|---|
| Maple-Preview | 20.2B / 1.49B | 75.1% | 87.5% | 78.8% | 73.5% | 78.7% |
| Qwen3.5 35B-A3B | 35B / 3B | 74.6% | 91.1% | 81.8% | 84.2% | 82.9% |
| GLM 4.7 Flash | 30B / 3B | 64.0% | 89.2% | 81.1% | 75.2% | 77.4% |
| Ternary Bonsai 27B (Qwen3.6 27B base) | 27.3B / — | 77.9% | 87.5% | 74.2% | 68.9% | 77.1% |
| Qwen3 30B-A3B | 30.5B / 3.3B | 66.0% | 88.3% | 78.8% | 73.4% | 76.6% |
| Qwen3.5 9B | 9B / — | 65.6% | 86.7% | 71.2% | 81.7% | 76.3% |
| GPT-OSS 20B | 20.9B / 3.6B | 74.6% | 90.0% | 68.9% | 71.5% | 76.3% |
| LFM2 24B-A2B | 24B / 2B | 21.9% | 23.2% | 17.4% | 47.0% | 27.4% |
Despite having by far the smallest active-parameter count (1.49B) among the comparison set, Maple-Preview's 78.7% average lands within a few points of every model except the larger Qwen3.5 35B-A3B — while being dramatically faster on-device thanks to its natively-ternary architecture. All comparison scores are as reported by DeepGrove in the same launch benchmark table; see each model's own page for scores from other sources.
FAQ
Is Maple-Preview open source? Yes — weights are released on Hugging Face under the MIT license, along with conversion tooling and an MLX-based inference fork for Apple Silicon.
How big is Maple-Preview? 20.2 billion total parameters with 1.49 billion active per token (20B-A1B mixture-of-experts), stored as 2-bit ternary weights in a 5.31 GB checkpoint.
What is Maple-Preview's context window? 131,072 tokens, using a hybrid 3:1 sliding-window (512 tokens) to global-attention design to bound KV-cache memory growth at long context.
How fast is Maple-Preview? 218 tokens/s on an Apple Mac mini M4, 281.5 tokens/s on a MacBook Pro (M5 Pro), and 127 tokens/s on an iPhone — 5–16x faster than comparably-capable models like Gemma 4, Qwen3.5, and gpt-oss, largely due to its natively-trained ternary-weight architecture.
What is Maple-Preview best used for? Math and coding reasoning tasks on-device, where decode speed and a small memory footprint matter most. DeepGrove explicitly notes this preview has had minimal agentic post-training, so it is not yet recommended for agentic workloads.
Benchmark scores sourced from DeepGrove's official Maple-Preview launch page (deepgrove.ai/maple-preview, Aug 4, 2026) and the deepgrove/maple-preview model card on Hugging Face. Comparison-model scores (Qwen3.5 35B-A3B, GLM 4.7 Flash, Ternary Bonsai 27B, Qwen3 30B-A3B, Qwen3.5 9B, GPT-OSS 20B, LFM2 24B-A2B) are as reported in that same table.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.