Part of Best Decision Model for Agent Routing and Scoring
Best Open-Weight Decision Model for Self-Hosting
Winnow-12B Q8 is the provisional first pick for Best Open-Weight Decision Model for Self-Hosting as of October 16, 2026, while independent task measurements are still limited.[1][2] It scores 68.89 on JevBench (#8 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 60 and its calibration axis 83, each out of 100. Cygnet (blockbrain, frozen Gemma-4-12B-it) scores within two points, so read the top two as close rather than as a measured gap. Cygnet (blockbrain, frozen Gemma-4-12B-it) is the next-ranked alternative. It scores 68.55 on JevBench (#9 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 55 and its calibration axis 87, each out of 100.
About this recommendation
- Updated
- Evidence through
- Sources
- 8
- Revision
- v18
Comparing Jev alternatives or choosing a System One model? Read our guide to AI decision models, routing, classification and scoring for the decision model vs LLM tradeoffs and a self-hosting checklist.
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
24
live candidates
Evaluation feeds
1
task-weighted
Winner coverage
100%
intended feed weight
Largest provider share
2 of 8
blockbrain
Winnow-12B Q8 leads the measured task comparison with limited independent task measurements. Community volume, release age and usage popularity add no ranking points. The leading scores are close under this policy.
Real-world track record is still developing. The recommendation can lead on independent measurements while practitioner reports accumulate.
Sources evaluated
These are the weights used in scoring. Missing measurements are omitted from each model’s average. A feed with no result for any model released in the last 60 days is stale and does not score until it catches up.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| JevBenchObserved 2026-10-16 · evaluation date not supplied · Winnow-12B Q8 · standard error 1.02 | 100% | #8 | 24/24 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- blockbrain2 models
- abhishek-rai1 model
- akhilaaa31 model
- eldanring1 model
- igor-l.-ninfer-contributors1 model
- mapika1 model
- us1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
eldanring · coverage 100% · provisional evidence
#8 JevBench
blockbrain · coverage 100% · close score · provisional evidence
#9 JevBench
akhilaaa3 · coverage 100% · close score · provisional evidence
#10 JevBench
abhishek-rai · coverage 100% · provisional evidence
#24 JevBench
blockbrain · coverage 100% · provisional evidence
#25 JevBench
igor-l.-ninfer-contributors · coverage 100% · provisional evidence
#28 JevBench
us · coverage 100% · provisional evidence
#29 JevBench
mapika · coverage 100% · provisional evidence
#30 JevBench
Task score averages the available, task-weighted measurements. Benchmark positions compare models matched to our catalog. Models measured by at least two independent evaluator families are considered first. Missing tests, practitioner report counts, popularity and release age add no points. Coverage describes evidence breadth, not a probability of correctness. Scores within two points are shown as close under this policy; this is not a statistical confidence interval. When the leading scores are that close, the model measured by more evaluator families ranks first.
- 01
Winnow-12B Q8eldanring · open-weight
Winnow-12B Q8 scores 68.89 on JevBench (#8 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 60 and its calibration axis 83, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 02
Cygnet (blockbrain, frozen Gemma-4-12B-it)blockbrain · open-weight
Cygnet (blockbrain, frozen Gemma-4-12B-it) scores 68.55 on JevBench (#9 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 55 and its calibration axis 87, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 03
Jev-Omni (akhilaaa3, Gemma-4-12B merged)akhilaaa3 · open-weight
Jev-Omni (akhilaaa3, Gemma-4-12B merged) scores 67.75 on JevBench (#10 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 56 and its calibration axis 87, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 04
spark-s1-4b-v6 (Open Spark Jev, abhishek085)abhishek-rai · open-weight
spark-s1-4b-v6 (Open Spark Jev, abhishek085) scores 44.9 on JevBench (#24 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 43 and its calibration axis 65, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 05
swanOne (blockbrain, Qwen3.8-Flash-Next NVFP4)blockbrain · open-weight
swanOne (blockbrain, Qwen3.8-Flash-Next NVFP4) scores 43.79 on JevBench (#25 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 56 and its calibration axis 89, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 06
NInfer Qwen3.8-Flash-Next mixedigor-l.-ninfer-contributors · open-weight
NInfer Qwen3.8-Flash-Next mixed scores 41.97 on JevBench (#28 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 49 and its calibration axis 90, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 07
jev-local (Qwen3.5-9B)us · open-weight
jev-local (Qwen3.5-9B) scores 41.33 on JevBench (#29 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 42 and its calibration axis 73, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
- 08
decider-4b v2 (Mapika)mapika · open-weight
decider-4b v2 (Mapika) scores 41.24 on JevBench (#30 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 40 and its calibration axis 90, each out of 100.
Best when: Its strongest result for this task is on JevBench, which measures calibrated typed decisions.No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.
Frequently asked
- What is the best model for Best Open-Weight Decision Model for Self-Hosting?
- Winnow-12B Q8 ranks first as of October 16, 2026. Winnow-12B Q8 scores 68.89 on JevBench (#8 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 60 and its calibration axis 83, each out of 100.[1]
- What is the best alternative to Winnow-12B Q8?
- Cygnet (blockbrain, frozen Gemma-4-12B-it) ranks second. Cygnet (blockbrain, frozen Gemma-4-12B-it) scores 68.55 on JevBench (#9 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 55 and its calibration axis 87, each out of 100.[2]
- Can I self-host the top pick for Best Open-Weight Decision Model for Self-Hosting?
- Yes. Winnow-12B Q8 is open-weight, so it can run on your own hardware as well as through hosted routes.[1]
Sources
- 1
“Within the 114 catalog models matched to this evaluation: Scores 68.89 on JevBench (#8 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 60 and its calibration axis 83, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 2
“Within the 114 catalog models matched to this evaluation: Scores 68.55 on JevBench (#9 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 55 and its calibration axis 87, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 3
“Within the 114 catalog models matched to this evaluation: Scores 67.75 on JevBench (#10 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 56 and its calibration axis 87, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 4
“Within the 114 catalog models matched to this evaluation: Scores 44.9 on JevBench (#24 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 43 and its calibration axis 65, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 5
“Within the 114 catalog models matched to this evaluation: Scores 43.79 on JevBench (#25 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 56 and its calibration axis 89, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 6
“Within the 114 catalog models matched to this evaluation: Scores 41.97 on JevBench (#28 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 49 and its calibration axis 90, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 7
“Within the 114 catalog models matched to this evaluation: Scores 41.33 on JevBench (#29 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 42 and its calibration axis 73, each out of 100.”
JevBench · Benchmark · Oct 16, 2026 - 8
“Within the 114 catalog models matched to this evaluation: Scores 41.24 on JevBench (#30 of 114), Benchmark Heaven's composite of decision accuracy, calibration, speed, and cost for typed decision models; its decision accuracy axis is 40 and its calibration axis 90, each out of 100.”
JevBench · Benchmark · Oct 16, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.