---

base_model: StandardThinking/StandardOne-3B

base_model_relation: quantized

license: apache-2.0

library_name: gguf

pipeline_tag: text-generation

tags:

- gguf

- llama.cpp

- mistral3

- decision-model

---

Standard One 3B — GGUF

> Updated weights (v2.2, 2026-10-04). These files are built from Standard One 3B v2.2.

> If you downloaded them before, download them again. Earlier releases remain available at v1, v1.1, v2, and v2.1.

> Decision API update (2026-10-08). All nine StandardOne-3B-*.gguf language model files now support stock llama.cpp's text /v1/systemone endpoint. Only five metadata fields were added; every tensor byte is unchanged. Download the updated files from main. The pre-update binaries remain at v2.2. mmproj is unchanged. See Decision API quick start.

Version: v2.2 weights + decision metadata (2026-10-08)

GGUF builds of Standard One 3B (Ministral 3 3B text + Pixtral vision tower,

mistral3 architecture) for use with llama.cpp.

Files

Most quant levels below (marked "imatrix") used an importance matrix calibrated on 1,512 prompts from this release's training rows. Q8_0 and BF16 do not use it.

| File | Quant | Size | imatrix |

|---|---|---:|:---:|

| StandardOne-3B-BF16.gguf | BF16 (no quantization) | 6.9 GB | — |

| StandardOne-3B-Q8_0.gguf | Q8_0 | 3.7 GB | no |

| StandardOne-3B-Q5_K_M.gguf | Q5_K_M | 2.5 GB | yes |

| StandardOne-3B-Q4_K_M.gguf | Q4_K_M | 2.1 GB | no |

| StandardOne-3B-IQ4_XS.gguf | IQ4_XS | 2.0 GB | yes |

| StandardOne-3B-Q3_K_M.gguf | Q3_K_M | 1.8 GB | yes |

| StandardOne-3B-IQ3_M.gguf | IQ3_M | 1.7 GB | yes |

| StandardOne-3B-Q2_K.gguf | Q2_K | 1.5 GB | yes |

| StandardOne-3B-IQ2_M.gguf | IQ2_M | 1.3 GB | yes |

| mmproj-StandardOne-3B.gguf | F16 vision projector | 840 MB | — |

SHA256 checksums: SHA256SUMS. Source revisions, conversion tool version and full validation

numbers: release-manifest.json.

Q4_K_M note: as in v2.1, the shipped file is the plain (no imatrix) build. For v2.1 both candidates were compared on the mean accuracy of 10 decision suites (6,836 questions); plain scored 69.53% and imat scored 69.33%.

Decision API quick start

Use any of the nine language model files in this repository; Q4_K_M is used below as an example.

The update uses llama.cpp's existing openjev letter-logit readout with a Standard One native template.

It is still Standard One weights. No engine fork, adapter server, or new weight conversion is needed.

To use another precision, replace the Q4_K_M filename in the download, checksum verification, and server commands with the selected filename from the table. All nine files use the same endpoint.

1. Install and download

Use the tested official llama.cpp b11495 release

for your platform. Extract the runtime, retain its companion libraries, and put

llama-server on PATH (or use the full executable path). This release reports

0.6.0-dev, commit 37ac63456. Older builds may lack the decision endpoint;

newer builds were not tested here.

In a Python 3.10+ virtual environment, install the Hub CLI:

python -m pip install --upgrade huggingface_hub
hf download StandardThinking/StandardOne-3B-GGUF \
  StandardOne-3B-Q4_K_M.gguf SHA256SUMS release-manifest.json \
  --revision main --local-dir models/standardone-3b-systemone

These shell examples use Bash/Zsh. The repository is public; login is unnecessary.

For a reproducible snapshot, replace main with a full commit hash from

Files and versions.

The v2.2 tag contains the original binaries without decision metadata. Use main or the metadata-update commit for this endpoint.

Verify just the downloaded model (the other entries in SHA256SUMS are optional):

python - <<'PYVERIFY'
import hashlib
from pathlib import Path
root = Path("models/standardone-3b-systemone")
name = "StandardOne-3B-Q4_K_M.gguf"
expected = {}
for line in (root / "SHA256SUMS").read_text().splitlines():
    if line.strip():
        digest, filename = line.split(maxsplit=1)
        expected[filename.lstrip("*")] = digest
h = hashlib.sha256()
with (root / name).open("rb") as stream:
    for chunk in iter(lambda: stream.read(8 * 1024 * 1024), b""):
        h.update(chunk)
if h.hexdigest() != expected.get(name):
    raise SystemExit("SHA256 mismatch")
print("SHA256 OK")
PYVERIFY

2. Run the local decision endpoint

llama-server \
  -m models/standardone-3b-systemone/StandardOne-3B-Q4_K_M.gguf \
  -ngl 0 -t 4 -tb 4 -c 4096 -np 1 --jinja \
  --host 127.0.0.1 --port 8080 --alias standardone-3b

This is the tested CPU configuration. The model file is about 2.1 GB; RAM use

also includes the KV cache and working buffers. Keep the server running. In

another terminal, wait until curl --fail http://127.0.0.1:8080/health returns

{"status":"ok"}. Stop the server with Ctrl+C when finished.

GPU offload and image input were not validated for this decision API.

3. Send Jev-style typed questions

curl --fail http://127.0.0.1:8080/v1/systemone \
  -H 'Content-Type: application/json' \
  --data-binary @- <<'JSON'
{
  "state": "A parcel has a large tear and its contents are visible.",
  "questions": {
    "route": {
      "type": "choice",
      "instructions": "Should this parcel pass inspection?",
      "criteria": {
        "pass": "The packaging is intact.",
        "review": "The packaging is visibly damaged.",
        "unclear": "There is not enough information."
      }
    },
    "damaged": {
      "type": "noul",
      "instructions": "Is the packaging visibly damaged?",
      "criteria": {
        "true": "A large tear is visible.",
        "false": "The packaging is intact."
      }
    },
    "damage": {
      "type": "score",
      "instructions": "How damaged is the packaging?",
      "criteria": ["intact", "minor cosmetic damage", "large tear with visible contents"]
    }
  }
}
JSON

Read answers.route.choice and answers.route.probabilities,

answers.damaged.noul (probability of yes), and answers.damage.score

(expected zero-based level). score can be fractional. The server constructs

JSON from option logits; the model does not generate a letter or JSON text.

usage.output_tokens is 0. A downloadable request is in

systemone-request.json.

The request demonstrates the response fields. Values differ across model sizes and quantizations; inspect the live response rather than treating one fixture as an accuracy guarantee.

What was verified, and what differs from the merged model's served endpoint

Each of the nine files passed six real-model fixtures: choice, noul, score, object state, reversed lexical choice-key order, and all three types in one request. Direct option-logprob readout agreed within 3.42e-08 for single-question fixtures and 3.42e-08 for the multi-question fixture. Multi-question reference calls follow the same question order and reuse prefix state after the first question. Independently prefilled references can differ due to CPU batching and cache behavior. Per-file results and tolerances are in systemone-validation.json. These checks verify endpoint calculations, not task accuracy.

• This API uses the native prompt, with temperatures choice=0.90,

noul=1.15, score=1.20. They are the published native preset carried forward

from v2.1 fitting, not temperatures fitted on these quantized weights.

• The current merged v2.2 default endpoint uses served wording, temperature

1.65, and extended uppercase labels. This implementation does not reproduce that preset.

• Supply at most 26 options. The embedded template rejects more; do not infer

Standard One support for the underlying openjev path's larger label count.

• Text state, string instructions, string choice descriptions, and

string score levels are the tested path. Object state is supported by

this template, but preserves submitted key order; the existing adapter's native

state renderer sorts keys. Keep state and option ordering fixed for comparisons.

• llama.cpp's confidence formulas differ from the existing adapter's normalized

entropy. Matching probability fields does not imply identical confidence semantics.

• All nine 3B language files were tested on CPU with text requests. GPU execution, images, structured question/description values, and complete request-option/error parity are unverified. Do not send images to this text-only template.

The template is published as systemone-native-template.jinja.

The release manifest records every added metadata field

and the identical tensor-data hash. Current checksums reflect the added metadata. Original binaries and checksums remain in the v2.2 snapshot; the manifest also records every source-file SHA256.

For the default BF16 endpoint, use the

merged model's serving instructions.

Importance-matrix calibration

The importance matrix used 1,512 prompts (up to 22 per cohort across 72 training-data cohorts), seed 20260924 and context 2048. The prompts came from this release's training rows.

Historical weight accuracy validation

The following measurements predate the metadata update. Tensor data is unchanged; these are not a new benchmark of the native decision endpoint.

Checked on CPU with llama-server from the same llama.cpp commit (no GPU offload), one question at a time, against the BF16 safetensors checkpoint run with transformers on the same token ids. The questions use native wording, no system prompt, one option order and at most 9 options: JevBench public easy and standard, judge proxy, realistic transfer and hard proxy (320 questions), or only the two JevBench tiers (120 questions). Accuracy is the most probable option.

| File | Questions | Accuracy | BF16 safetensors | Same answer as BF16 | Mean total-variation distance |

|---|---:|---:|---:|---:|---:|

| BF16 | 320 | 75.62 % | 75.31 % | 98.75 % | 0.006 |

| Q8_0 | 320 | 75.62 % | 75.31 % | 98.12 % | 0.010 |

| Q5_K_M | 120 | 93.33 % | 94.17 % | 97.50 % | 0.015 |

| Q4_K_M | 320 | 75.00 % | 75.31 % | 93.12 % | 0.045 |

| IQ4_XS | 120 | 95.00 % | 94.17 % | 98.33 % | 0.028 |

| Q3_K_M | 120 | 95.83 % | 94.17 % | 98.33 % | 0.045 |

| IQ3_M | 120 | 93.33 % | 94.17 % | 95.83 % | 0.051 |

| Q2_K | 120 | 91.67 % | 94.17 % | 90.83 % | 0.113 |

| IQ2_M | 320 | 70.94 % | 75.31 % | 77.50 % | 0.155 |

Note on the mmproj conversion

llama.cpp's stock --mmproj converter (as of the commit used here) drops the [IMG_BREAK] token

embedding for HF-format Mistral3ForConditionalGeneration checkpoints (a filter meant to strip

text-model tensors also strips the one row of the text embedding matrix the vision projector needs),

so the mmproj file it produces fails to load in llama-server/llama-cli

("unable to find tensor v.token_embd.img_break"). The mmproj file in this folder was built with a

small local patch that lets that one tensor through; see release-manifest.json ->

known_issues_fixed for details. It loads and runs correctly with --mmproj.

Update history

2026-10-08 — llama.cpp decision API metadata

• Added metadata for stock llama.cpp's /v1/systemone endpoint to all nine language GGUF files: BF16, Q8_0, Q5_K_M, Q4_K_M, IQ4_XS, Q3_K_M, IQ3_M, Q2_K, and IQ2_M.
• Added the decision type, a named native decision template, and per-type temperatures. The v2.2 weights and every tensor byte are unchanged; the vision projector (mmproj) is unchanged.
• Verified choice, noul, and score with official llama.cpp b11495 on every precision: 54 API fixture requests for this repository, including object state, option order, and multi-question requests. See systemone-validation.json.
• Added decision endpoint instructions and request examples, and updated SHA256SUMS and release-manifest.json. This implementation uses native wording with at most 26 options and text input; it does not reproduce the merged model's default served preset.
• Pre-update GGUF binaries remain available at v2.2. Metadata update commit.

License

Apache License 2.0 — see LICENSE and NOTICE. Same terms as the source StandardOne-3B release;

this GGUF conversion adds no additional restrictions.