Updated weights (v2, 2026-09-26). These files are built from Standard One 3B v2. If you downloaded them before, download them again or
pin revision="v2". Earlier versions stay available under the tags v1 and v1.1.
Version: v2
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") were built with an importance matrix calibrated on 1,512
prompts sampled from our own training rows (see "Importance-matrix calibration" below), which
recovers some of the accuracy quantization would otherwise lose. Q8_0 and BF16 don't need one.
| 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 |
yes |
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: this is the imatrix-calibrated version, not a plain quantization. We generated both and
chose whichever scored higher on the mean of 10 held-out and public-dataset decision suites (no JevBench items; imatrix 69.77 vs. plain 69.63).
Usage
Text-only:
llama-cli -m StandardOne-3B-Q4_K_M.gguf -ngl 99 -p "Your prompt"
With vision (image input):
llama-server -m StandardOne-3B-Q4_K_M.gguf --mmproj mmproj-StandardOne-3B.gguf -ngl 99
The GGUF's own embedded chat template (converted from the model's chat_template.jinja) is applied
automatically; no extra flags needed for chat formatting.
Importance-matrix calibration
Q5_K_M down to IQ2_M were quantized with llama-imatrix calibrated on 1,512 prompts (42 per cohort across 36 training-data cohorts; training data only — no benchmark/held-out file was used), context 2048. Q8_0 and BF16 don't
use an imatrix (high enough precision that it doesn't move the needle).
Validation
Accuracy was checked by comparing next-token logits over the option letters on the JevBench public
suites (easy/original/hard) against the served BF16 baseline; see release-manifest.json for full
methodology and gguf-validation.md (in the release kit) for the complete writeup. Measured
(accuracy %, n=48/72/111 for easy/original/hard; prefill tokens/sec is the mean over the 231 scored
decisions):
| Quant |
Easy |
Original |
Hard |
Overall |
Prefill tok/s |
| served BF16 (reference) |
100.0 |
88.89 |
46.85 |
— |
— |
| BF16-GGUF |
100.0 |
87.50 |
51.35 |
72.73 |
11,672 |
| Q8_0 |
100.0 |
87.50 |
51.35 |
72.73 |
7,842 |
| Q5_K_M |
100.0 |
88.89 |
46.85 |
71.00 |
4,966 |
| Q4_K_M (shipped, imatrix) |
100.0 |
88.89 |
45.95 |
70.56 |
6,770 |
| IQ4_XS |
100.0 |
90.28 |
52.25 |
74.03 |
6,181 |
| Q3_K_M |
100.0 |
87.50 |
47.75 |
71.00 |
6,565 |
| IQ3_M |
100.0 |
83.33 |
45.95 |
68.83 |
6,632 |
| Q2_K |
100.0 |
80.56 |
40.54 |
65.37 |
4,774 |
| IQ2_M |
97.92 |
70.83 |
49.55 |
66.23 |
5,082 |
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.
License
Apache License 2.0 — see LICENSE and NOTICE. Same terms as the source StandardOne-3B release;
this GGUF conversion adds no additional restrictions.
