The standard way to make a large model small is quantization: train at full precision, then crush the weights into lower-bit formats after the fact. It works, but there’s a structural tension — the model spent training learning to use float32 or bfloat16 values, and you’re forcing it to approximate those values with far fewer bits once the training is over.
Maple-Preview, released today by DeepGrove, takes the other route. Each weight is constrained to the ternary set {-1, 0, +1} throughout training, not imposed afterward. That’s roughly 1.58 bits of information per parameter — below INT2, and substantially below the 4-bit quantization most optimized runtimes use. The optimizer never has access to higher-precision weights to begin with, which forces it to find solutions that are naturally representable in ternary rather than approximations of solutions that weren’t.
Stack that on a Mixture-of-Experts architecture — 24 layers, 256 experts, 8 active per token — and the effective parameter count per forward pass is around 1 billion out of 20 billion total. The resulting checkpoint is 5.31 GB. On an M4 Mac mini, Maple-Preview runs at 218 tokens per second. On an iPhone, it reaches 120 tokens per second.
To put those numbers in context: the Swiftlet runtime covered here yesterday achieved 4.5–5 tok/s for an 80B model by streaming expert weights from SSD on demand. Maple-Preview’s approach is structurally different — it’s not about smarter IO; it’s about making the model itself smaller at the representation level, so it fits into memory without special tricks. Swiftlet and Maple-Preview aren’t competing solutions so much as complementary ones: one optimizes the access pattern, the other optimizes the weight density.
The attention layout is a 3:1 ratio of sliding-window attention (SWA-512) to global attention layers, with a 131K token context. Sliding-window attention caps the per-token attention span to 512 positions in most layers, which significantly reduces the memory cost of the KV cache over long contexts; the periodic global attention layers let the model attend across the full sequence when needed. This is a well-established design pattern for long-context efficiency, and it’s sensible here given the memory constraints.
The benchmark claims are the more surprising part. The model reports competitive scores on AIME 2026 and HMMT 2026 — math competitions that test hard problem-solving, not pattern matching — as well as GPQA-D, which evaluates graduate-level science reasoning. DeepGrove describes it as a new point on the Pareto frontier for memory-to-performance and speed-to-performance in its weight class.
The model card is honest about the current state: Maple-Preview “received minimal post-training for agentic tasks and only small-scale general reinforcement learning.” The Hacker News comments surfaced this quickly — the model is prone to confident-sounding but factually incorrect responses on knowledge retrieval tasks. The reasoning scaffolding is there; the general-purpose reliability isn’t fully developed yet. This is a preview.
DeepGrove previously published Bonsai, a ternary 0.5B model trained from scratch earlier in 2026. Maple-Preview applies the same training methodology at 40× the scale, layered with MoE. The fact that the approach is scaling is the noteworthy result — ternary training isn’t a hack that works only for tiny models. The Spectra paper from 2024 was an early demonstration of this, showing that ternary LLMs trained from scratch can be surprisingly competitive at scale; DeepGrove is extending that finding into MoE territory.
MIT license, weights on Hugging Face. A ternary-weight MoE with competitive reasoning benchmarks, in 5.31 GB, running at native speed on consumer Apple hardware, is a reasonable base for experimentation. The “preview” label signals that the interesting part — post-training on agentic tasks, broader RL — is still ahead.