FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

Suite · grounding← Suites

Grounding and output mode

Does the mitigation work, and what does each output mode cost?

Cases
125
Scored runs
61
Models
60
Top score
100.0%
SpecificationMethod · grading · caveat
Method
The same question is asked of the same document under different output modes, including free text and constrained decoding. Documents carry line addresses so any claim can be traced back to a source line.
Grading
A case passes when the answer is present in the document and correctly attributed. Answers taken from the wrong line count as misattribution, not as fabrication.
Caveat
Requires a server that honours response_format. Results from mlx_lm are not valid because it ignores the field.
Version
grounding/1
Cases
125
Recorded runs61 runs · sorted by accuracy

Sorted by accuracy. The bar shows the 95 percent confidence interval. Where intervals overlap, the models are not distinguishable on this suite at this number of cases.

Top of the suite Qwen3-Next-80B-A3B Thinking 100.0% 95% interval 20.7–100.0

ModelScore95% interval PassedRuntime · profileRecorded
Qwen3-Next-80B-A3B Thinkingqwen3-next-80b-thinking100.0%20.7–100.0Passed 1/1llama.cppstudy-default2026-08-27
qwen3.8-27b-mlx-4bitunenforced92.0%85.9–95.6overlaps the top intervalPassed 115/125mlxextraction2026-09-07
edge0-35b89.6%83.0–93.8overlaps the top intervalPassed 112/125llama.cppextraction2026-09-16
Qwen3-Next-80B-A3B Instructqwen3-next-80b88.8%82.1–93.2overlaps the top intervalPassed 111/125llama.cppextraction2026-09-05
qwen3-next-80b-a3b-iq4xs88.8%82.1–93.2overlaps the top intervalPassed 111/125llama.cppextraction2026-08-26
qwen3-next-80b-mlx-4bitunenforced87.2%80.2–92.0overlaps the top intervalPassed 109/125mlxextraction2026-09-07
Qwen3-Next-80B-A3B Instruct, REAP prunedqwen3-next-80b-reap87.2%80.2–92.0overlaps the top intervalPassed 109/125llama.cppextraction2026-09-08
granite-4.2-30b-mlx-q4unenforced86.4%79.3–91.3overlaps the top intervalPassed 108/125mlxextraction2026-09-08
Muse-Glimmer 30Bmuse-glimmer-30b85.6%78.4–90.7overlaps the top intervalPassed 107/125llama.cppextraction2026-09-08
Gemma 4 12Bgemma4-12b84.0%76.6–89.4overlaps the top intervalPassed 105/125llama.cppextraction2026-08-24
gemma4-12b-b1061884.0%76.6–89.4overlaps the top intervalPassed 105/125llama.cppextraction2026-08-25
Granite 4.2 30Bgranite-4.2-30b-gguf84.0%76.6–89.4overlaps the top intervalPassed 105/125llama.cppextraction2026-09-08
Qwen3.6 35B-A3Bqwen3.6-35b82.4%74.8–88.1overlaps the top intervalPassed 103/125llama.cppextraction2026-08-27
flashnext-4.27bpw-m6481.6%73.9–87.4overlaps the top intervalPassed 102/125llama.cppextraction2026-09-08
ornith-1.5-9b81.6%73.9–87.4overlaps the top intervalPassed 102/125llama.cppextraction2026-09-12
Qwen3.5 122B-A10Bqwen3.5-122b-a10b-iq380.8%73.0–86.7overlaps the top intervalPassed 101/125llama.cppextraction2026-08-26
ornith-9b-kvq480.0%72.1–86.1overlaps the top intervalPassed 100/125llama.cppextraction2026-09-12
Qwen3.8 27Bqwen3.8-27b80.0%72.1–86.1overlaps the top intervalPassed 100/125llama.cppextraction2026-08-24
qwen3.8-27b-noreason80.0%72.1–86.1overlaps the top intervalPassed 100/125llama.cppextraction2026-09-07
Devstral Small 2 24Bdevstral-small-2-24b-b1061879.2%71.3–85.4overlaps the top intervalPassed 99/125llama.cppextraction2026-08-25
Qwen3.8-Flash-Nextqwen38-flash-next79.2%71.3–85.4overlaps the top intervalPassed 99/125llama.cppextraction2026-09-05
Qwen3 Coder 30Bqwen3-coder-30b77.6%69.5–84.0overlaps the top intervalPassed 97/125llama.cppextraction2026-08-24
BTL-4 Compact (IQ2_XXS)btl-476.8%68.7–83.3overlaps the top intervalPassed 96/125llama.cppextraction2026-09-15
BTL-4 (Q4_K_M)btl-4-q4km76.8%68.7–83.3overlaps the top intervalPassed 96/125llama.cppextraction2026-09-15
glm-4.7-flash-mlx-q8unenforced76.8%68.7–83.3overlaps the top intervalPassed 96/125mlxextraction2026-09-08
qwen3-coder-30b-b1061876.8%68.7–83.3overlaps the top intervalPassed 96/125llama.cppextraction2026-08-26
qwen3-next-80b-coder76.8%68.7–83.3overlaps the top intervalPassed 96/125llama.cppextraction2026-08-27
GLM-4.7-Flashglm-4.7-flash-q876.0%67.8–82.6overlaps the top intervalPassed 95/125llama.cppextraction2026-08-26
ornith-35b-a3b-q4076.0%67.8–82.6overlaps the top intervalPassed 95/125llama.cppextraction2026-09-12
qwen3-8b-mlx-8bitunenforced74.4%66.1–81.2overlaps the top intervalPassed 93/125mlxextraction2026-09-08
qwen3.5-4b-base74.4%66.1–81.2overlaps the top intervalPassed 93/125llama.cppextraction2026-09-13
Ornith 1.5 35Bornith-1.5-35b-q6-b1061872.8%64.4–79.8overlaps the top intervalPassed 91/125llama.cppextraction2026-08-25
qwen3.5-9b-q872.0%63.6–79.1overlaps the top intervalPassed 90/125llama.cppextraction2026-08-24
qwen3.5-9b-q8-b1061872.0%63.6–79.1overlaps the top intervalPassed 90/125llama.cppextraction2026-08-25
Agents-A1 35Bagents-a1-35b71.2%62.7–78.4overlaps the top intervalPassed 89/125llama.cppextraction2026-08-24
agents-a1-35b-b1061871.2%62.7–78.4overlaps the top intervalPassed 89/125llama.cppextraction2026-08-26
bonsai-27b-mlx2-cortextunenforced71.2%62.7–78.4overlaps the top intervalPassed 89/125mlxextraction2026-09-14
bonsai-27b-mlx2-m4prounenforced71.2%62.7–78.4overlaps the top intervalPassed 89/125mlxextraction2026-09-14
Ternary-Bonsai-27B (prism-ml)ternary-bonsai-27b68.8%60.2–76.3overlaps the top intervalPassed 86/125llama.cppextraction2026-09-14
qwen3-next-80b-a3b-iq4xs-smoke68.0%48.4–82.8overlaps the top intervalPassed 17/25llama.cppextraction—
Bonsai-27B (Q1_0)bonsai-27b-1bit67.2%58.6–74.8overlaps the top intervalPassed 84/125llama.cppextraction2026-09-14
ornith-1.5-35b-mlx-4bitunenforced67.2%58.6–74.8overlaps the top intervalPassed 84/125mlxextraction2026-09-08
Qwen3.8-9B-Distill (Q8_0)qwen3.8-9b-distill67.2%58.6–74.8overlaps the top intervalPassed 84/125llama.cppextraction2026-09-14
bonsai2-27b-pq266.4%57.7–74.1overlaps the top intervalPassed 83/125llama.cppextraction2026-09-18
bonsai2-27b-ptq166.4%57.7–74.1overlaps the top intervalPassed 83/125llama.cppextraction2026-09-18
Qwen3-Next-80B-A3B Thinkingqwen3-next-80b-thinking65.6%56.9–73.4overlaps the top intervalPassed 82/125llama.cppextraction2026-08-27
neohorse-1-4b63.2%54.5–71.1overlaps the top intervalPassed 79/125llama.cppextraction2026-09-13
gemma4-e4b62.4%53.7–70.4overlaps the top intervalPassed 78/125llama.cppextraction2026-08-24
gemma4-e4b-b1061862.4%53.7–70.4overlaps the top intervalPassed 78/125llama.cppextraction2026-08-25
minicpm5-2b62.4%53.7–70.4overlaps the top intervalPassed 78/125llama.cppextraction2026-09-15
Gemma 4 E2Bgemma4-e2b61.6%52.8–69.7overlaps the top intervalPassed 77/125llama.cppextraction2026-08-24
gemma4-e2b-b1061861.6%52.8–69.7overlaps the top intervalPassed 77/125llama.cppextraction2026-08-25
Nemotron 3.5 Lightning 30Bnemotron-lightning-30b60.0%51.2–68.2overlaps the top intervalPassed 75/125llama.cppextraction2026-09-05
nemotron-lightning-30b-mlx-4bitunenforced56.8%48.0–65.2overlaps the top intervalPassed 71/125mlxextraction2026-09-08
minicpm5-2b-q8-np156.0%47.2–64.4overlaps the top intervalPassed 70/125llama.cppextraction2026-09-17
Bonsai-8B (Q1_0)bonsai-8b-1bit55.2%46.5–63.6overlaps the top intervalPassed 69/125llama.cppextraction2026-09-14
k2-horizon-7b55.2%46.5–63.6overlaps the top intervalPassed 69/125llama.cppextraction2026-09-09
glm-4.7-flash-mlx-q4unenforced53.6%44.9–62.1overlaps the top intervalPassed 67/125mlxextraction2026-09-08
ternary-bonsai-8b51.2%42.5–59.8overlaps the top intervalPassed 64/125llama.cppextraction2026-09-17
minicpm5-2b-exl348.0%39.4–56.7overlaps the top intervalPassed 60/125llama.cppextraction—
LFM 2.5 8B-A1Blfm2.5-8b-a1b47.2%38.7–55.9overlaps the top intervalPassed 59/125llama.cppextraction2026-08-31

Runs marked unenforced were measured through mlx_lm on a suite that requires constrained decoding. mlx_lm accepts the schema and ignores it, so those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.

Run log63 logged runs of this suite
Every logged run, newest first (63)
FinishedModelRuntime · profileScoreCoverageGrader · dataset · config
bonsai2-27b-ptq1not recordedextraction66.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 0407c0f9d876e4a3
bonsai2-27b-pq2not recordedextraction66.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 1685634443e1889f
minicpm5-2b-q8-np1not recordedextraction56.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 5ac3f164bce7795b
minicpm5-2b-exl3not recordedextraction60.0%Coverage 80% coverageg 30b933dcc6a7e2d4d dcb396faf8b83b0ac deed1276a964eb4d
ternary-bonsai-8bnot recordedextraction51.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac a964482c0a6fefb9
edge0-35bnot recordedextraction89.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac eb1910ec33534645
minicpm5-2bnot recordedextraction62.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 8f32a36bd07ac12d
btl-4-q4kmnot recordedextraction76.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac b6421e71dc1b07d5
btl-4not recordedextraction76.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 7d9f8293fd874d62
bonsai-27b-1bitnot recordedextraction67.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 0f1cdb00af0adb60
bonsai-8b-1bitnot recordedextraction55.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 1ac5570cbae78ecd
bonsai-27b-mlx2-cortextmlxextraction71.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 4f595a317a905809
bonsai-27b-mlx2-m4promlxextraction71.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 86fb3d4c7b477c6c
ternary-bonsai-27bnot recordedextraction68.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 3b0afe72fb0c0121
qwen3.8-9b-distillnot recordedextraction67.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 696fd8b30eed53f0
qwen3.5-4b-basenot recordedextraction74.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 0f7dbcd0b1d1afd4
neohorse-1-4bnot recordedextraction63.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac b382340e2e3ad6fa
ornith-9b-kvq4not recordedextraction80.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 4c61ed44aed10e7a
ornith-35b-a3b-q40not recordedextraction76.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 95c69b1144e96db8
ornith-1.5-9bnot recordedextraction81.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 4a6ad96362484d18
k2-horizon-7bnot recordedextraction55.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 30ddb902b06a0c00
flashnext-4.27bpw-m64not recordedextraction81.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 2867d56b04f8bbb4
qwen3-next-80b-reapnot recordedextraction87.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 7e1bbf58551d87ef
muse-glimmer-30bllama.cppextraction85.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac d9d25da365e04cc8
glm-4.7-flash-mlx-q4mlxextraction53.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 2432689cc6751eea
glm-4.7-flash-mlx-q8mlxextraction76.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 6d0a345ab4ecfa54
qwen3-8b-mlx-8bitmlxextraction74.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac e34a89daf5181d58
nemotron-lightning-30b-mlx-4bitmlxextraction56.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 15a43d7b2024f68c
granite-4.2-30b-ggufllama.cppextraction84.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac e353f603986dcdcd
ornith-1.5-35b-mlx-4bitmlxextraction67.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac a438aff9707ec8a1
granite-4.2-30b-mlx-q4mlxextraction86.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac db0a9174aa469c68
qwen3-next-80b-mlx-4bitmlxextraction87.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 0c539d39afff5bd1
qwen3.8-27b-noreasonllama.cppextraction80.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 4a3b4b7c53219329
qwen3.8-27b-mlx-4bitmlxextraction92.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 42a3cf291ff302bb
qwen38-flash-nextllama.cppextraction79.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 9148fcdc2bc14fe4
nemotron-lightning-30bllama.cppextraction60.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac f734b41b3caabfdd
qwen3-next-80bllama.cppextraction88.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac f34a40087b318cc6
lfm2.5-8b-a1bllama.cppextraction47.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 90be8dffe935ab41
qwen3-next-80b-coderllama.cppextraction76.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 140e26c8be639be7
qwen3.6-35bllama.cppextraction82.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 279f3a4bd1ff0c7e
qwen3.6-35bllama.cppextraction86.3%Coverage 73/73g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 279f3a4bd1ff0c7e
qwen3-next-80b-thinkingllama.cppstudy-default100.0%Coverage 1/1g 30b933dcc6a7e2d4d dcb396faf8b83b0ac cb1f0b0a42b6ac48
qwen3-next-80b-thinkingllama.cppextraction65.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac e644b21297e6198b
qwen3-next-80bllama.cppextraction88.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac f34a40087b318cc6
qwen3-next-80b-a3b-iq4xsllama.cppextraction88.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac e3b83bfa70c1a417
qwen3-next-80b-a3b-iq4xs-smokellama.cppextraction73.9%Coverage 92% coverageg 30b933dcc6a7e2d4d dcb396faf8b83b0ac e6920f17163814d7
glm-4.7-flash-q8llama.cppextraction76.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac f7116f1eb6a7f001
qwen3.5-122b-a10b-iq3llama.cppextraction80.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac b3059734754cd76c
agents-a1-35b-b10618llama.cppextraction71.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac b6d15c86cef9c0f6
qwen3-coder-30b-b10618llama.cppextraction76.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 85522d05c3671c16
qwen3.5-9b-q8-b10618llama.cppextraction72.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 9e54265c33ff0f82
gemma4-12b-b10618llama.cppextraction84.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac ba97a71cfa0468a7
gemma4-e2b-b10618llama.cppextraction61.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac b18ad353c84d839e
gemma4-e4b-b10618llama.cppextraction62.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac dc6988d619b065bf
ornith-1.5-35b-q6-b10618llama.cppextraction72.8%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac bc2f7777482c6fb6
devstral-small-2-24b-b10618llama.cppextraction79.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac ed588aa2abdb41ef
qwen3-coder-30bllama.cppextraction77.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac ade06bde17b71832
qwen3.8-27bllama.cppextraction80.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 1d989a6b00eb212c
agents-a1-35bllama.cppextraction71.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 3cb35299a047efb8
qwen3.5-9b-q8llama.cppextraction72.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac bd21ebb86ee6f863
gemma4-12bllama.cppextraction84.0%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 31241ed46ed4d6a5
gemma4-e4bllama.cppextraction62.4%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 267624907124ed7e
gemma4-e2bllama.cppextraction61.6%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac dee67653e2de45c6

Open in the run log →