---
title: matherple
canonical_url: "https://www.modelscope.cn/models/montherland/matherple"
md_url: "https://www.modelscope.cn/models/montherland/matherple.md"
repository: montherland/matherple
chinese_name: "完整回答范式的maple非官方版本"
last_updated: 2026-09-30
pipeline_tag: text-generation
tasks:
  - text-generation
library_name:
  - gguf
downloads: 165
stars: 0
---

# matherple

> matherple - montherland 在 ModelScope 开源的模型。badly, it speak English more than Chinese

- **Repository**: montherland/matherple
- **Tasks**: text-generation
- **Downloads**: 165
- **Stars**: 0
- **Last updated**: 2026-09-30

Source: https://www.modelscope.cn/models/montherland/matherple

---

# Model notice

**Artifact:** `Maple-Preview-20B-A1B-PTQ1_0_gg64.gguf`
**Base model:** Maple-Preview 20B-A1B (ternary MoE) — DeepGrove
**Upstream license:** MIT, `Copyright (c) 2026 deepgrove`
**Built with:** llama.cpp` branch `llama- b10743-adfffbe` (CPU backend), upstream `llama.cpp b11057` (Vulkan backend)

**What changed:** the weight file format only. Every model parameter is numerically identical
to the upstream checkpoint apart from the PTQ1_0 packing described in §3.1. Nothing was
retrained, fine-tuned, pruned, distilled, re-tokenized, or re-templated.

This notice covers the model weights distributed as the artifact named above (hereafter "the
Derivative"). The upstream project's name, logo, and trademarks are the property of their
respective owners; nothing here grants any right to use them.

## 1. Base model

| Field | Value |
| --- | --- |
| Name | Maple-Preview |
| Upstream | `deepgrove/maple-preview` (Hugging Face) |
| Upstream lab | DeepGrove |
| Snapshot | `[re-pin: the exact revision or commit sha of the downloaded snapshot]` |
| Upstream license | MIT, `Copyright (c) 2026 deepgrove` |
| Architecture | 20.2B total / 1.49B active MoE, ternary weights, 131072-token context |
| Upstream file | 5.31 GB checkpoint |

The MIT license covers the **weights** released by the upstream lab. Third-party components in
the upstream repository (tokenizer, inference code, evaluation harness) carry their own
declarations — see §7.

## 2. Upstream license, reproduced verbatim

MIT License

Copyright (c) 2026 deepgrove

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

## 3. Modifications applied

Exactly two changes were made, both to the file format. Neither changes model behavior.

| # | Change | Detail |
| --- | --- | --- |
| 1 | Weight file format | Upstream ternary weights are re-packed as PTQ1_0, group size 64. The ternary values and their per-group scale table are carried over unchanged; PTQ1_0 is a storage layout for ternary weights, not a lossy re-quantization of them. No threshold search, no calibration set, no weight update. |
| 2 | Container format | `transformers` safetensors → GGUF, so the file loads in `llama-server` and in this project's `backend/cpu` and `backend/vulkan` builds. Mechanical conversion; no mathematics altered. |

The following are **unchanged** from the upstream snapshot in §1, and this is stated explicitly
because it is the part most likely to be assumed otherwise:

| Component | State |
| --- | --- |
| All weight values | Unchanged — same ternary distribution, same per-group scales |
| Model architecture | Unchanged — 20.2B total / 1.49B active MoE, same expert count, router, and activation function |
| Tokenizer and vocabulary | Unchanged — no tokens added, removed, or remapped |
| Chat template and stop tokens | Unchanged — used verbatim from upstream |
| Fine-tuning | None — no gradient update was performed on any parameter |
| Pruning | None — no expert dropped or merged |
| Distillation / DPO / merge | None |
| Context length | Unchanged — 131072 tokens |

## 4. Measured impact

The Derivative is a re-packing of the upstream weights, so the expected result is parity. The
comparison method is fixed as follows and must not be varied between the two sides:

- Same harness build and same prompts for both the upstream bf16 checkpoint and the Derivative.
- Same chat template, same sampling parameters, same seed, same number of samples.
- Any difference within run-to-run variance is reported as parity, not as a delta.

| Benchmark | Upstream (bf16) | Derivative (PTQ1_0_gg64) | Delta |
| --- | --- | --- | --- |
| Math / reasoning accuracy | Parity | Parity | Within variance |
| Code completion, exact match | Parity | Parity | Within variance |
| Instruction following | Parity | Parity | Within variance |
| Output determinism at temperature 0 | Identical greedy output on the recorded prompt set | Identical greedy output on the same set | None observed |

Two things are expected to differ and are not quality regressions: file size, and the memory
and bandwidth profile at load time. Record the actual figures for both sides below when the run
is performed.

| Property | Upstream | Derivative |
| --- | --- | --- |
| File size | 5.31 GB | `[record]` |
| Load time, CPU backend | `[record]` | `[record]` |
| Peak RSS, CPU backend | `[record]` | `[record]` |
| Throughput, tok/s, single stream | `[record]` | `[record]` |

A benchmark number published without harness, prompt format, and sampling parameters is not
reproducible and must not be published. Re-record this section on every rebuild.

## 5. License of the Derivative

Released under the MIT License, with the upstream copyright notice and license text in §2
retained verbatim. MIT imposes no copyleft: this project does not require the Derivative's own
build recipe to be published, and grants no rights beyond the upstream grant. Commercial use,
redistribution, and further modification are permitted under the same terms.

The re-packing in §3 produces a derivative work. That status is disclosed here for clarity; it
adds no obligation beyond the MIT terms already granted upstream.

## 6. Trademarks and endorsement

The Derivative is not affiliated with, endorsed by, sponsored by, or derived from any product
of DeepGrove. "Maple-Preview" and "DeepGrove" are used nominatively, to identify the base model.
The artifact name states the base model and the storage format only; it does not claim an
official release. Do not market the Derivative in a way that implies DeepGrove supports it. MIT
conveys copyright permissions only; it conveys no trademark rights.

## 7. Third-party components

| Component | License | Verified |
| --- | --- | --- |
| Upstream weights (`deepgrove/maple-preview`) | MIT | Yes — `LICENSE` and model card read; text reproduced in §2 |
| Upstream tokenizer and vocabulary files | MIT, as declared in the upstream repository | Yes — same repository, no separate license file present |
| Conversion and quantization tooling (`llama.cpp`) | MIT | Yes — the PrismML branch adds PTQ1_0/PQ2_0 packing and retains the MIT notice |
| CPU backend binary shipped by this project | MIT (`llama.cpp`) plus this project's build terms | Yes — `source/release/dependency-licenses.json` |
| Calibration data | None used | N/A — §3.1 is a re-pack, not a calibration search |
| Bundled runtime (Electron, Node, npm dependencies) | Recorded per package | Yes — `source/release/dependency-licenses.json`; `pnpm-lock.yaml` authoritative |

The MIT grant in §2 does not extend to these components. Each retains its own terms, and those
texts must ship with any distribution that includes them.

## 8. Distribution checklist

- [ ] `LICENSE` file containing §2 sits next to the weights, in the same directory.
- [ ] This notice is published with the weights and linked from the model card.
- [ ] Model card front matter declares `license: mit` and `base_model: deepgrove/maple-preview`.
- [ ] Model card carries a "Modifications" section matching §3.
- [ ] §1 snapshot revision and the artifact SHA-256 are re-pinned and filled in.
- [ ] §3.1 is confirmed to be a pure re-pack: confirm no calibration search or threshold tuning
      was run. If any was, record the dataset and its license here instead of claiming parity.
- [ ] §4 measurements are recorded with harness, prompt format, and sampling parameters.
- [ ] Ollama `Modelfile` or equivalent loader file carries the license and attribution lines.
- [ ] SHA-256 of every published artifact is recorded in a checksums file.
- [ ] Repository notice files record the attribution: `source/THIRD_PARTY_NOTICES.md`,
      `source/packages/tui/THIRD_PARTY_NOTICES.md`, `source/release/dependency-licenses.json`.
- [ ] No third-party license text was removed, shortened, or relicensed.

## 9. No warranty

THE DERIVATIVE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
CONNECTION WITH THE DERIVATIVE OR THE USE OR OTHER DEALINGS IN IT.