Expand Nex-N2.5-mini GGUF ladder and fidelity card

README.md

@@ -75,17 +75,24 @@ mean PPL was `6.684743` in this pilot.
 | Nex-N2.5-mini-Q6_K.gguf | 0.027186 | 93.811% | -0.858% | 4.676% |
 | Nex-N2.5-mini-Q5_K_M.gguf | 0.038910 | 92.636% | +0.036% | 5.262% |
 | Nex-N2.5-mini-Q4_K_M.gguf | 0.063583 | 90.315% | +3.230% | 6.732% |
+| Nex-N2.5-mini-IQ4_NL.gguf | 0.062792 | 90.273% | +1.378% | 6.774% |
+| Nex-N2.5-mini-IQ4_XS.gguf | 0.065216 | 90.071% | +0.801% | 6.854% |
+| Nex-N2.5-mini-Q3_K_L.gguf | 0.122060 | 86.334% | +3.083% | 9.158% |
 | Nex-N2.5-mini-Q3_K_M.gguf | 0.129316 | 85.906% | +4.044% | 9.427% |
+| Nex-N2.5-mini-IQ3_M.gguf | 0.172758 | 83.735% | +13.676% | 11.555% |
+| Nex-N2.5-mini-IQ3_S.gguf | 0.154186 | 84.600% | +10.201% | 10.701% |
 | Nex-N2.5-mini-Q2_K.gguf | 0.233133 | 80.709% | +10.839% | 12.770% |
+| Nex-N2.5-mini-Q2_K_S.gguf | 0.278980 | 78.709% | +15.232% | 14.023% |
 | Nex-N2.5-mini-IQ2_XS.gguf | 0.516971 | 71.590% | +51.851% | 19.371% |
 | Nex-N2.5-mini-IQ1_M.gguf | 0.695877 | 66.064% | +69.116% | 24.275% |
 | Nex-N2.5-mini-Q1_0.gguf | 8.616836 | 4.879% | +445003.665% | 61.752% |
 
 For a general local profile, Q4_K_M is the practical starting point in this
-pilot. Q5_K_M and Q6_K are stronger quality/size choices, while Q8_0 is the
-highest-bit option. Q3_K_M is a lower-memory compromise. Q2_K, IQ2_XS, IQ1_M,
-and Q1_0 are memory-constrained experimental profiles and should be checked
-against the intended workload.
+pilot. IQ4_NL and IQ4_XS are compact Q4-region alternatives. Q5_K_M and Q6_K
+are stronger quality/size choices, while Q8_0 is the highest-bit option.
+Q3_K_L, Q3_K_M, IQ3_M, and IQ3_S are lower-memory Q3-region compromises.
+Q2_K, Q2_K_S, IQ2_XS, IQ1_M, and Q1_0 are memory-constrained experimental
+profiles and should be checked against the intended workload.
 
 These measurements describe next-token fidelity relative to BF16; they are not
 a direct percentage of capabilities retained. Instruction following,
@@ -136,7 +143,7 @@ upstream BF16 safetensors. The text converter used `--no-mtp` because this
 revision advertises MTP configuration but does not contain MTP tensors. The
 vision projector was converted separately to F16.
 
-All nine published text GGUF files passed llama.cpp tensor checks, load, and
+All fifteen published text GGUF files passed llama.cpp tensor checks, load, and
 English generation smoke tests. The BF16 reference also passed the same text
 smoke profile. The Q4_K_M text file and the included F16 projector passed a
 multimodal image smoke test. Runtime throughput is supplementary and is

SHA256SUMS.txt

@@ -1,25 +1,31 @@
 cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30  LICENSE
 e13761d3e82e9ae6c21c3c6d6c4a812ef19105a3310c09adfa1d407ceb0952cb  Nex-N2.5-mini-IQ1_M.gguf
 c4bc72337f51431262e22cc7cef92ae6b22bf788328c7fd4f360a47927393401  Nex-N2.5-mini-IQ2_XS.gguf
+fa8af1d3860ecad6aafc808357786c868f15b01681aeba11f5be9f84bfc19d20  Nex-N2.5-mini-IQ3_M.gguf
+8f0e0deda6ad7ff8fc4a4186d6e2a090bb57605f3dea3a2f06cc261ade358e44  Nex-N2.5-mini-IQ3_S.gguf
+48179f330ce195ec46005fdd17ce55a1d3617fb66d669c77879be731ab833f87  Nex-N2.5-mini-IQ4_NL.gguf
+ba9a9db4d80293b9603e469b224f2c54fddc3c07d005122f6c40aae4fba41ec1  Nex-N2.5-mini-IQ4_XS.gguf
 f5d0faeaa7d187d2bce9b2b06923ac5e24f9ad96e18613235095e95d36966790  Nex-N2.5-mini-Q1_0.gguf
 009f420aed015eee2e418a1ca68ab75b4a477c6f504a89993edcd58f5ea701c4  Nex-N2.5-mini-Q2_K.gguf
+00c81ec5d66751e0e50a24ce540bc3f19cdd26604435b79c83e4fda76b12a6bb  Nex-N2.5-mini-Q2_K_S.gguf
+0487558daca42d6fc2737386219364721c2025199802c86bcc8724ce1ff84fbb  Nex-N2.5-mini-Q3_K_L.gguf
 dc92fc762dd63eed44f945e0940476d9c1ae3f4a28babf3ec5d3a4b0da64c3a0  Nex-N2.5-mini-Q3_K_M.gguf
 1cae631a01539db4d3f559683ef1c659f61dcf261f9f63489750665a3380c449  Nex-N2.5-mini-Q4_K_M.gguf
 a8fafb845de52adae9c09dd65da211372a23c8194a1a498b4173c6b245ab0404  Nex-N2.5-mini-Q5_K_M.gguf
 9ec1715e749525879a9cee386a79f87f06c19e509544fb4e44fd1a46dbdd2c0a  Nex-N2.5-mini-Q6_K.gguf
 c51b8b6b1c03dd929afd2ce19b4c6b96ea26eecdfce5822ec0b4ae153fa10d93  Nex-N2.5-mini-Q8_0.gguf
-77c946f53ff035903dc33922fc90cb7d939ad956afb5aac68fa9de141a8243e8  README.md
+9f134129b8ff35cddae009a116ad03df7a37333fd21797666919f64e41c647d1  README.md
 007fdc6302c9be799e145629fc02062a527075a9d5610a0dd41139a5750003f5  assets/Nex-N2.5-Benchmark-white.png
 f1753536417ee87cded4bc5017354eb2123c079d20c634403a3d808b2ba3cc5b  chat_template.jinja
 c128ba00f092ecc75a15fbc31174588faac26cbfbc3d0bff4b828f73f69604e4  mmproj-Nex-N2.5-mini-F16.gguf
-20b5125daebc3e607cf11a2076764b2d2bf2b9e54fc132da53580617007f216e  reproducibility/README.md
-c81d34e651521832c169e0199fcd24c244446bec62483e00afbcf36f558036fe  reproducibility/benchmark_gpu0.sh
-32830fcfd062809e16f22b5a45f5844e00836a92bf9f26fefa008326458c4ced  reproducibility/gpu0_smoke_test_ladder.sh
-937a78374825d6038b279a77dd0e5c93df6cf5ee75f390440d69aecbaf2b9d1e  reproducibility/manifest.md
+cb0f092074135932a0ea52d525dad369dff862a409c2aab198f1239052065ae9  reproducibility/README.md
+9ad6df11ac209b8a1087d3248dc964ea5c5b81fea4ec852d68586d4550f2f5e0  reproducibility/benchmark_gpu0.sh
+a72faf14ffa464e929167a4ded723b6be3d17686e18bda08f31ecef0b209425c  reproducibility/gpu0_smoke_test_ladder.sh
+cf0a8529ca4afc6cc3f5e987bfcd2b908c35481295f59e1ba2b908b583a16e3b  reproducibility/manifest.md
 68b8ad6d5fdbb973c5728ce691c60e8812346ebf48256a968b7bf1c1514ad143  reproducibility/mmproj_smoke_test_gpu0.sh
-bb91ebfd90ed8ee04e8783baa77779eb4c8189e5587438cbd24a65d558b9d3e5  reproducibility/quality-summary.tsv
-50f042f9a4bfff977bbad458e65dc2de75c541f380182cba8fb7e22ece4d2906  reproducibility/quality_eval_gpu0.sh
-3d13281bdd8d317f68d996f287027739b04744874bf1bcd163d15029b8f04b0b  reproducibility/quantize_ladder.sh
-38b856bdec5ba97bc34c96cd2e9ddd39c1d602b571da53289af106a36f2527de  reproducibility/runtime-summary.tsv
+be778f892e6b692fe2d0a3efd1a1d8f884e1f2dcc28a06d2cb801994ad961cb6  reproducibility/quality-summary.tsv
+82786e6931b5e78e8d409ea6c9144e8377a7323e3be43929bc00d7d3ccd3c048  reproducibility/quality_eval_gpu0.sh
+900dd394278723ae0d8bfe45f25c2e2119b728d65bc072884e7f2f938a55625a  reproducibility/quantize_ladder.sh
+a5b892528fdcd57e70c8d16e22e202e3f87c64f2f2f7d25da8f4abf8fb3699bf  reproducibility/runtime-summary.tsv
 6df41caec8c89eb5270606d4d479c63872283cbd46647352bde0cb19a8a9b9f6  reproducibility/summarize_bench.py
-9fdd4fa33ede09269b23e7b586e3649b707a18918bf024248743f911ecc89f8f  reproducibility/summarize_fidelity.py
+9adb3adbbd62211e99410a5afb17007b1d6306f8f5b32f1a3759cc9175d6c809  reproducibility/summarize_fidelity.py

reproducibility/README.md

@@ -1,9 +1,10 @@
 # Reproduction notes
 
 The release files were produced from the locked Nex-N2.5-mini BF16 GGUF input,
-not from another quantized file. The local workspace paths are supplied as
-defaults; set `WORKSPACE_ROOT`, `SOURCE`, `IMATRIX`, and `QUANTIZER` when
-reproducing elsewhere.
+not from another quantized file. The default ladder contains 15 profiles; set
+`QUANTS` to a space-separated subset when running one or more profiles. The
+local workspace paths are supplied as defaults; set `WORKSPACE_ROOT`, `SOURCE`,
+`IMATRIX`, and `QUANTIZER` when reproducing elsewhere.
 
 ```bash
 bash reproducibility/quantize_ladder.sh
@@ -12,6 +13,13 @@ bash reproducibility/mmproj_smoke_test_gpu0.sh
 bash reproducibility/benchmark_gpu0.sh
 ```
 
+For example, to reproduce only selected profiles:
+
+```bash
+QUANTS="IQ4_NL IQ3_M Q2_K_S" bash reproducibility/quantize_ladder.sh
+QUANTS="IQ4_NL IQ3_M Q2_K_S" bash reproducibility/gpu0_smoke_test_ladder.sh
+```
+
 The smoke, multimodal, and benchmark scripts pin `CUDA_VISIBLE_DEVICES=0` and
 use `CUDA0`. The benchmark is a runtime profile only; it is not a quality
 score. Raw logs remain in local staging under `reports/` and are not part of

reproducibility/benchmark_gpu0.sh

@@ -7,6 +7,12 @@ WORKSPACE_ROOT="${WORKSPACE_ROOT:-$(cd -- "$PACKAGE_ROOT/.." && pwd)}"
 BENCH="${BENCH:-$WORKSPACE_ROOT/tools-llama.cpp-qwen4exp-mtp-direct/build-a10m-direct/bin/llama-bench}"
 OUTDIR="${OUTDIR:-$WORKSPACE_ROOT/reports/nex-n2.5-mini/benchmark-gpu0}"
 SOURCE="${SOURCE:-$WORKSPACE_ROOT/source-official-bf16-nex-n2.5-mini/Nex-N2.5-mini-BF16.gguf}"
+ALL_MODELS=(BF16 Q8_0 Q6_K Q5_K_M Q4_K_M IQ4_NL IQ4_XS Q3_K_L Q3_K_M IQ3_M IQ3_S Q2_K Q2_K_S IQ2_XS IQ1_M Q1_0)
+if [[ -n "${QUANTS:-}" ]]; then
+  read -r -a models <<< "$QUANTS"
+else
+  models=("${ALL_MODELS[@]}")
+fi
 
 [[ -x "$BENCH" ]] || { echo "Missing BENCH: $BENCH" >&2; exit 1; }
 [[ -s "$SOURCE" ]] || { echo "Missing BF16 SOURCE: $SOURCE" >&2; exit 1; }
@@ -25,7 +31,6 @@ export CUDA_VISIBLE_DEVICES=0
   nvidia-smi --query-gpu=index,name,memory.total,memory.used,utilization.gpu --format=csv
 } > "$OUTDIR/environment.txt"
 
-models=(BF16 Q8_0 Q6_K Q5_K_M Q4_K_M Q3_K_M Q2_K IQ2_XS IQ1_M Q1_0)
 for quant in "${models[@]}"; do
   if [[ "$quant" == BF16 ]]; then
     model="$SOURCE"

reproducibility/gpu0_smoke_test_ladder.sh

@@ -7,6 +7,12 @@ WORKSPACE_ROOT="${WORKSPACE_ROOT:-$(cd -- "$PACKAGE_ROOT/.." && pwd)}"
 CLI="${CLI:-$WORKSPACE_ROOT/tools-llama.cpp-qwen4exp-mtp-direct/build-a10m-direct/bin/llama-cli}"
 OUTDIR="${OUTDIR:-$WORKSPACE_ROOT/reports/nex-n2.5-mini/smoke-tests-gpu0}"
 THREADS="${THREADS:-64}"
+ALL_MODELS=(Q8_0 Q6_K Q5_K_M Q4_K_M IQ4_NL IQ4_XS Q3_K_L Q3_K_M IQ3_M IQ3_S Q2_K Q2_K_S IQ2_XS IQ1_M Q1_0)
+if [[ -n "${QUANTS:-}" ]]; then
+  read -r -a models <<< "$QUANTS"
+else
+  models=("${ALL_MODELS[@]}")
+fi
 
 [[ -x "$CLI" ]] || { echo "Missing CLI: $CLI" >&2; exit 1; }
 mkdir -p "$OUTDIR"
@@ -17,7 +23,6 @@ export CUDA_VISIBLE_DEVICES=0
 summary="$OUTDIR/summary.tsv"
 printf 'quantization\tstatus\tlog\n' > "$summary"
 
-models=(Q8_0 Q6_K Q5_K_M Q4_K_M Q3_K_M Q2_K IQ2_XS IQ1_M Q1_0)
 for quant in "${models[@]}"; do
   model="$PACKAGE_ROOT/Nex-N2.5-mini-${quant}.gguf"
   log="$OUTDIR/${quant}.log"

reproducibility/manifest.md

@@ -75,8 +75,9 @@ published quantization was produced directly from the BF16 GGUF by
 `reproducibility/quantize_ladder.sh` with 128 quantization threads:
 
 - `Q8_0`: direct BF16 quantization without an imatrix
-- `Q6_K`, `Q5_K_M`, `Q4_K_M`, `Q3_K_M`, `Q2_K`, `IQ2_XS`, `IQ1_M`, `Q1_0`:
-  BF16 source plus the model-specific imatrix
+- `Q6_K`, `Q5_K_M`, `Q4_K_M`, `IQ4_NL`, `IQ4_XS`, `Q3_K_L`, `Q3_K_M`,
+  `IQ3_M`, `IQ3_S`, `Q2_K`, `Q2_K_S`, `IQ2_XS`, `IQ1_M`, `Q1_0`: BF16 source
+  plus the model-specific imatrix
 
 No published file was used as the source for another quantization.
 
@@ -87,7 +88,7 @@ the test processes. The exact executable and flags are retained in the public
 reproduction scripts, while raw logs remain local under `reports/`.
 
 - BF16 reference: text tensor check, load, and English generation pass
-- 9/9 text GGUF artifacts: tensor check, load, and English generation pass
+- 15/15 text GGUF artifacts: tensor check, load, and English generation pass
 - `Nex-N2.5-mini-Q4_K_M.gguf` plus
   `mmproj-Nex-N2.5-mini-F16.gguf`: image load and English description pass
 - Text smoke profile: 2,048 context, 64 generated tokens, full GPU-layer
@@ -115,7 +116,7 @@ difference. The raw reference-logit files and evaluation logs are local only.
 ## Runtime profile
 
 `reproducibility/runtime-summary.tsv` contains the compact runtime comparison
-for BF16 and the nine quantizations. It was measured with `llama-bench` using
+for BF16 and the fifteen quantizations. It was measured with `llama-bench` using
 256 prompt tokens, 64 generation tokens, batch/ubatch 512, 64 CPU threads,
 three repetitions, full GPU-layer setting, `--fit-target 1024`, and one CPU MoE
 layer setting. Runtime throughput is supplementary and should not be read as a

reproducibility/quality-summary.tsv

@@ -3,8 +3,14 @@ Q8_0	0.023277	94.523	6.750630	6.684743	0.986	4.132
 Q6_K	0.027186	93.811	6.627382	6.684743	-0.858	4.676
 Q5_K_M	0.038910	92.636	6.687126	6.684743	0.036	5.262
 Q4_K_M	0.063583	90.315	6.900634	6.684743	3.230	6.732
+IQ4_NL	0.062792	90.273	6.776839	6.684743	1.378	6.774
+IQ4_XS	0.065216	90.071	6.738274	6.684743	0.801	6.854
+Q3_K_L	0.122060	86.334	6.890803	6.684743	3.083	9.158
 Q3_K_M	0.129316	85.906	6.955046	6.684743	4.044	9.427
+IQ3_M	0.172758	83.735	7.598980	6.684743	13.676	11.555
+IQ3_S	0.154186	84.600	7.366668	6.684743	10.201	10.701
 Q2_K	0.233133	80.709	7.409299	6.684743	10.839	12.770
+Q2_K_S	0.278980	78.709	7.702984	6.684743	15.232	14.023
 IQ2_XS	0.516971	71.590	10.150869	6.684743	51.851	19.371
 IQ1_M	0.695877	66.064	11.304952	6.684743	69.116	24.275
 Q1_0	8.616836	4.879	29754.033896	6.684743	445003.665	61.752

reproducibility/quality_eval_gpu0.sh

@@ -10,6 +10,12 @@ TEST="${TEST:-$WORKSPACE_ROOT/calibration/wikitext-2-raw/wiki.test.raw}"
 VALID="${VALID:-$WORKSPACE_ROOT/calibration/wikitext-2-raw/wiki.valid.raw}"
 OUTDIR="${OUTDIR:-$WORKSPACE_ROOT/reports/nex-n2.5-mini/quality-eval-gpu0}"
 CHUNKS="${CHUNKS:-8}"
+ALL_QUANTS=(Q8_0 Q6_K Q5_K_M Q4_K_M IQ4_NL IQ4_XS Q3_K_L Q3_K_M IQ3_M IQ3_S Q2_K Q2_K_S IQ2_XS IQ1_M Q1_0)
+if [[ -n "${QUANTS:-}" ]]; then
+  read -r -a QUANT_LIST <<< "$QUANTS"
+else
+  QUANT_LIST=("${ALL_QUANTS[@]}")
+fi
 
 [[ -x "$PPL" ]] || { echo "Missing PPL binary: $PPL" >&2; exit 1; }
 [[ -s "$BF16" && -s "$TEST" && -s "$VALID" ]] || { echo "Missing BF16 or hold-out corpus" >&2; exit 1; }
@@ -39,7 +45,7 @@ for split in test valid; do
   fi
 done
 
-for quant in Q8_0 Q6_K Q5_K_M Q4_K_M Q3_K_M Q2_K IQ2_XS IQ1_M Q1_0; do
+for quant in "${QUANT_LIST[@]}"; do
   model="$PACKAGE_ROOT/Nex-N2.5-mini-${quant}.gguf"
   [[ -s "$model" ]] || { echo "Missing model: $model" >&2; exit 1; }
   for split in test valid; do
@@ -54,4 +60,5 @@ done
 
 python3 "$SCRIPT_DIR/summarize_fidelity.py" "$OUTDIR" \
   "$PACKAGE_ROOT/reproducibility/quality-summary.tsv" \
-  --chunks "$CHUNKS"
+  --chunks "$CHUNKS" \
+  --quants "${QUANT_LIST[@]}"

reproducibility/quantize_ladder.sh

@@ -11,6 +11,12 @@ QUANTIZER="${QUANTIZER:-$WORKSPACE_ROOT/tools-llama.cpp-qwen4exp-mtp-direct/buil
 OUTDIR="${OUTDIR:-$PACKAGE_ROOT}"
 LOGDIR="${LOGDIR:-$WORKSPACE_ROOT/reports/nex-n2.5-mini/quantize-logs}"
 THREADS="${THREADS:-128}"
+ALL_QUANTS=(Q8_0 Q6_K Q5_K_M Q4_K_M IQ4_NL IQ4_XS Q3_K_L Q3_K_M IQ3_M IQ3_S Q2_K Q2_K_S IQ2_XS IQ1_M Q1_0)
+if [[ -n "${QUANTS:-}" ]]; then
+  read -r -a QUANT_LIST <<< "$QUANTS"
+else
+  QUANT_LIST=("${ALL_QUANTS[@]}")
+fi
 
 [[ -s "$SOURCE" ]] || { echo "Missing SOURCE: $SOURCE" >&2; exit 1; }
 [[ -x "$QUANTIZER" ]] || { echo "Missing QUANTIZER: $QUANTIZER" >&2; exit 1; }
@@ -37,13 +43,8 @@ run_quant() {
   echo "done $type"
 }
 
-# Standard release ladder. Every output is produced directly from SOURCE.
-run_quant Q8_0
-run_quant Q6_K
-run_quant Q5_K_M
-run_quant Q4_K_M
-run_quant Q3_K_M
-run_quant Q2_K
-run_quant IQ2_XS
-run_quant IQ1_M
-run_quant Q1_0
+# Every output is produced directly from SOURCE. QUANTS can narrow a run, e.g.
+# QUANTS="IQ4_NL IQ4_XS" bash quantize_ladder.sh
+for quant in "${QUANT_LIST[@]}"; do
+  run_quant "$quant"
+done

reproducibility/runtime-summary.tsv

@@ -2,8 +2,14 @@ model	file_size_bytes	file_size_gib	relative_to_bf16_pct	prompt_tok_s	generation
 Nex-N2.5-mini-BF16.gguf	69376636992	64.61	100.0	81.49	15.58	3
 Nex-N2.5-mini-IQ1_M.gguf	8239208480	7.67	11.9	708.86	113.61	3
 Nex-N2.5-mini-IQ2_XS.gguf	10507030560	9.79	15.1	1306.48	107.88	3
+Nex-N2.5-mini-IQ3_M.gguf	15429530112	14.37	22.2	1409.79	103.02	3
+Nex-N2.5-mini-IQ3_S.gguf	15239434752	14.19	22.0	1410.51	101.16	3
+Nex-N2.5-mini-IQ4_NL.gguf	19768289792	18.41	28.5	952.33	77.28	3
+Nex-N2.5-mini-IQ4_XS.gguf	18717788672	17.43	27.0	1265.77	83.57	3
 Nex-N2.5-mini-Q1_0.gguf	5316528160	4.95	7.7	2556.30	141.15	3
 Nex-N2.5-mini-Q2_K.gguf	12939593760	12.05	18.7	1193.68	110.57	3
+Nex-N2.5-mini-Q2_K_S.gguf	12141107712	11.31	17.5	1171.30	120.28	3
+Nex-N2.5-mini-Q3_K_L.gguf	18104340992	16.86	26.1	1341.48	97.17	3
 Nex-N2.5-mini-Q3_K_M.gguf	16764764192	15.61	24.2	1338.77	100.31	3
 Nex-N2.5-mini-Q4_K_M.gguf	21166757920	19.71	30.5	702.23	78.02	3
 Nex-N2.5-mini-Q5_K_M.gguf	24729131040	23.03	35.6	438.27	53.74	3

reproducibility/summarize_fidelity.py

@@ -4,6 +4,7 @@
 from __future__ import annotations
 
 import argparse
+import os
 import re
 from pathlib import Path
 
@@ -36,9 +37,12 @@ def main() -> None:
     parser.add_argument("logdir", type=Path)
     parser.add_argument("output", type=Path)
     parser.add_argument("--chunks", type=int, default=8)
+    parser.add_argument("--quants", nargs="+", help="quantization names to summarize")
     args = parser.parse_args()
 
-    quants = ("Q8_0", "Q6_K", "Q5_K_M", "Q4_K_M", "Q3_K_M", "Q2_K", "IQ2_XS", "IQ1_M", "Q1_0")
+    all_quants = ("Q8_0", "Q6_K", "Q5_K_M", "Q4_K_M", "IQ4_NL", "IQ4_XS", "Q3_K_L", "Q3_K_M", "IQ3_M", "IQ3_S", "Q2_K", "Q2_K_S", "IQ2_XS", "IQ1_M", "Q1_0")
+    env_quants = os.environ.get("QUANTS", "").split()
+    quants = tuple(args.quants or env_quants or all_quants)
     rows: list[tuple[str, dict[str, float]]] = []
     for quant in quants:
         records = [