Home / Models / Maple-Preview 20B-A1B

How to Run Maple-Preview 20B-A1B Locally — Setup Guide (September 2026)

Maple-Preview is a 20B reasoning Mixture-of-Experts model whose weights are ternary — each one packed to roughly 1.6 bits as −α, 0 or +α. The result is a 20B model that lands under 6 GB on disk, small enough for an 8 GB card to hold with a long context. Like the other ternary models, its kernels do not exist in stock llama.cpp, so it is a build-from-source job.

Numbers you can trust

We don't print guessed speeds. Every measured number on this page came from TurboLLM's own auto-tuner running Maple-Preview 20B-A1B on real hardware — labeled as such. When you run it yourself, TurboLLM benchmarks it on your exact GPU and shows a VRAM-fit verdict before you load, and the real measured tokens/sec once it's running.

Does it fit your GPU?

20.2B total parameters across 256 experts (8 active, ~1B active per token), MIT-licensed. 24 layers with a 3:1 pattern of 512-token sliding-window attention to full attention, giving a very small KV cache. Native 131K context.

Maple-Preview (ternary)
DeepGrove · GGUF
MoE · ~1B active · ternary weights
thinkingcode
VRAMFit
8 GBTQ1_0-head-Q4_K (5.0 GB) or TQ2_0-head-Q4_K (5.9 GB) fully in VRAM · 131K context
12 GB+TQ2_0-head-F16 (6.4 GB) — the best of the ternary set — with a long context · 131K context
deepgrove/maple-preview-GGUF
View on Hugging Face →

How to run it

  1. Install TurboLLM

    npx turbollm — no install step, works on Windows, macOS, and Linux. It detects your GPU and auto-provisions a matching llama-server build (no CUDA toolkit, no Python, no compiler).

  2. Get the model

    Paste deepgrove/maple-preview-GGUF into TurboLLM's in-app Hugging Face search — it lists every quant with a VRAM-fit verdict against your real free VRAM, and downloads are resumable and SHA-256 verified.

  3. Let auto-fit pick the setup

    Flip the auto-fit toggle and TurboLLM chooses the GPU/CPU layer split (and, for a Mixture-of-Experts model, the expert-offload split) for your card — the same manual flag-hunting every other guide walks you through by hand.

  4. Load it

    TurboLLM benchmarks the model on your exact GPU as it loads and shows the real measured tokens/sec — not a number copied from someone else's card.

  5. Use it

    Chat in the built-in UI, or point any OpenAI- or Anthropic-compatible tool (including Claude Code) at TurboLLM's local API — see the API docs.

Need the DeepGrove llama.cpp fork instead?

The maple architecture is not in mainline llama.cpp — a search of the upstream repository returns nothing for it — so a stock engine cannot load these files at any quant. DeepGrove publish their own MIT-licensed build at https://github.com/deepgrove-ai/llama.cpp; paste that URL into TurboLLM's "Add via git repo" flow and it compiles with CUDA in one pass. Note that the Hugging Face repo's "Use with llama.cpp" snippet is auto-generated boilerplate, not a claim of stock support — ignore it.

One honest caveat

This is a preview release and its own model card says so: it "received minimal post-training for agentic tasks and only small-scale general reinforcement learning". Treat it as an interesting look at where ternary quantization is going rather than a daily driver — and note that every speed claim published for it so far is first-party. We have not measured this one ourselves yet, so there are deliberately no numbers on this page.

FAQ

Will Maple-Preview 20B-A1B run on my GPU?

Check the VRAM table above for the tier closest to your card. If you're not sure which tier you're in, or want picks tailored to your exact hardware, use What can I run? — enter your VRAM (or Apple unified memory) and it suggests the model, quant, and offload split that fits, the same logic TurboLLM's auto-fit runs when you load a model.

Which quant should I use for Maple-Preview 20B-A1B?

Pick the highest quant that fits your VRAM with room left over for the KV cache and compute buffer — not just the model weights. As a starting point, Q4_K_M is usually the sweet spot (small quality loss for a large size drop); step up to Q5_K_M/Q6_K/Q8_0 if you have headroom, or down to Q3_K_M/an IQ quant if you don't. TurboLLM shows a VRAM-fit verdict for every quant against your card's real free VRAM before you download — see Quantization explained for what the letters and numbers actually mean, or llama.cpp flags that actually matter for the other settings worth knowing.

Run it now

$ npx turbollm

One command detects your GPU, provisions the right engine, and opens the UI. New here? Start with Install & first run and Quantization explained. On different hardware, see RTX 4060 or RTX 5060 Ti. Also see how to run it or how to run it.