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.cppbranchllama- b10743-adfffbe(CPU backend), upstreamllama.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
- [ ]
LICENSEfile 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: mitandbase_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
Modelfileor 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.