Downloads · 30 days
3.6K
35% of all-time downloads
SZLHOLDINGS/SZL-Khipu-1.5B-GGUF
SZL-Khipu-1.5B-GGUF is a text generation model from SZLHOLDINGS. Use it when you need the model to write or continue text. It is set up for llama.cpp. The card lists the license as apache-2.0.
<p<a href="https://huggingface.co/spaces/SZLHOLDINGS/szl-command-lab"<img src="https://raw.githubusercontent.com/szl-holdings/szl-brand/27d12e77fc0f9f1a0e5fae8b34adf2380d51d070/kit/logos/orbit-v2/suite/szl-icon-32.svg…
Downloads · 30 days
3.6K
35% of all-time downloads
All-time downloads
10.4K
Public
Repo size
6.9 GB
Likes
1
Public
Click a slice to open those files.
.gguf6.9 GB · 100%
From the Hugging Face model README
Compact Khipu quantizations for research on retrieval-plan proposals over controller-supplied handles.
Artifact: GGUF quantizations · Stage: Research · promotion blocked
Explore in Command Lab · Build · Evidence
Compact, proposal-only retrieval plans over supplied handles
Derived quantizations · limited research/demo scope · production promotion blocked
GGUF derivatives of SZLHOLDINGS/SZL-Khipu-1.5B, itself a fine-tune of Qwen/Qwen2.5-1.5B-Instruct. The model proposes a JSON retrieval plan over candidate handles supplied by a controller. It does not contain or authorize access to a private graph, execute the plan, or issue trusted receipts.
Reviewed Hub revision: 7c39154b22ccb5e2151b4dd53e0d36965dfbc920, inspected 2026-09-30 UTC.
The following exact file identities are recorded in publication.json. File presence was checked against the reviewed tree; this review did not download or independently hash the GGUF bytes.
| File | Bytes recorded | SHA-256 recorded |
|---|---|---|
SZL-Khipu-1.5B-F16.gguf | 3093668768 | 2348ee342efe639e100f3fb31a3dc11b8c12d8c43ecfe45e18041b9f94c71a12 |
SZL-Khipu-1.5B-Q4_K_M.gguf | 986047904 | 13c1a1993063e1dff92f7413ccf48eaca6d48efc8801ae9af35961ae3396623a |
SZL-Khipu-1.5B-Q5_K_M.gguf | 1125049760 | 3bf460ac163c5dc952c273999c38a41349e3e6d666e4b713aed22c996860fd4c |
SZL-Khipu-1.5B-Q8_0.gguf | 1646572448 | 6aff1087f64631679f4cdf032613aee6911dbde38cd3bac6b81bf63741a56f0d |
File size does not equal peak runtime memory. This card makes no ranking of quantization quality, speed or energy efficiency because no comparative post-quantization evaluation was established.
The repository carries owner-signed training/evaluation receipts from the pre-quantized model and a repository-declared public key. The publication record reports that these signatures validated against that key; this review did not repeat signature verification. These are Ed25519 signatures over canonical JSON, not DSSE envelopes. External signer identity is not independently established.
The recorded pre-quantized held-out evaluation has plan-validity 11/11, grounding 4/5 and abstention 2/6. Those results belong to the evaluated artifact described in the receipt. They do not establish behavioral performance for any of these GGUF files. Keep the abstention failure visible.
An unsigned publication record also describes a historical best-effort CPU service probe with two runs and a median latency of 1467.5 ms. That tiny service observation is not a current uptime assertion, an SLA, a benchmark, restart reproducibility, or a post-quantization quality test. Instrumented energy and process-peak memory are unavailable in that record. Current service operation was not tested in this card review.
The source of this mirror's card and banner is the separate Forge publisher input
khipu-gguf/card/README.md. A card-only correction does not attest the GGUF
build, publish weights, or qualify a runtime.
publication.json identifies the source curriculum, schema and receipt tree at szl-forge@cddba1ad9887211d6fd87386e98bb6707f16dbf5/khipu. It records upstream base revision 989aa7980e4cf806f80c7fef2b1adb7bc71aa306 and fine-tuned parent revision 759c6112b1acae67f30351ef4e652e07671a42fb.
This identifies a source relationship; it does not establish a reproducible quantization build. No owner-signed release receipt covering these GGUF bytes is reported. A parent-model signature does not cover a derived GGUF unless that exact file digest is included in its signed subject.
Use a reviewed, pinned llama.cpp-compatible runtime and the exact selected file. The prompt contract is khipu.schema.json: a query plus candidate handles, followed by a proposed NAVIGATE or ABSTAIN plan. Validate output schema and handle membership independently. The controller retains access, approval and execution authority.
An immutable artifact retrieval example, not an inference or deployment test:
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="SZLHOLDINGS/SZL-Khipu-1.5B-GGUF",
filename="SZL-Khipu-1.5B-Q4_K_M.gguf",
revision="7c39154b22ccb5e2151b4dd53e0d36965dfbc920",
)
print(path)
Retrieval may consume approximately 986 MB of storage for the selected file plus cache overhead. Hash-check against the reviewed evidence before use. This example was not executed in the review; confirm disk and runtime memory headroom before running it. A downloaded file is not a validated model.
Artwork: SZL Orbit · © 2026 SZL Holdings / Stephen Lutar · CC BY 4.0.
| Surface | Link |
|---|---|
| Company | szlholdings.com |
| Sponsors | Become a sponsor |
| GitHub | szl-holdings |
| Receipts chain | frontier-54-runs |
| Live dashboard | SI command center |