Solar Decide

+ contribute

Solar Decide is a closed-weights decision model from Upstage, released 22 September 2026. Licence: proprietary.

Catalogue identity notes

Named native decision endpoint verified 2026-10-31. Medical study records exact requested and returned model IDs and pinned Upstage routing. Historical dated service snapshot retained in each result; no claim that current served weights are immutable.

Model information sources

Release date ↗

56.5±1.4#144 of 355 decision modelsfrom 32 benchmark results across 3 sources

$ how it compares

The decision models ranked either side of Solar Decide, out of 357 with a score, then more from Upstage.

#	model	score	where	released
144	CoCo-Decision-4B 2.4 Corners	57±1	model files	2026-10-16
145	Amber Hybrid 4B (decision readout) MLukauskas	57±1	model files
146	Solar Decide Upstage	57±1	api price unknown	2026-09-22
147	Sieve-Vision-4B Sthānika AI	56±4	model files
148	xor-26b-a4b-v1.1 Juspay Technologies	56±4	model files	2026-10-09
	more from Upstage
223	Solar Decide Flash Upstage	48±1	api price unknown

see all decision models · find the ones that fit your machine

$ what it's good at

retrieval & classification▁▁▁·▁▁▁▁▃▃█▅ #179 of 258 · 20% of its score · 8 tests

Trails Mola-68M, Bosun v3.1 1.7B LoRA, Dinah-1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
System One Bench: MIRACL pairwise (Arabic)	85.5% ✎	▂···▁··▃▄▂█▇ beats 6/29	easy · avg 84%	▇▇ 4.8%	at its level
System One Bench: MASSIVE intent (Arabic)	75.0% ✎	▄·▁·▁▃▃▃█▄▇▇ beats 5/29	easy · avg 76%	▇▇ 4.5%	typical
LeanMCP · scifact rank noul	55.7% ✎	▃····▂···▃▅█ beats 0/11	medium · avg 62%	▇ 3.5%	at its level
System One Bench: MIRACL rerank (Arabic)	84.5% ✎	▁▂·▁▁·▂·▄▄█▇ beats 10/29	easy · avg 79%	▇ 2.8%	typical
LeanMCP · nfcorpus rank noul	32.4% ✎	▃▃·····▃▃▅██ beats 0/11	hard · avg 29%	▇ 2.5%	blunt · at its level
System One Bench: MASSIVE scenario (Arabic)	55.5% ✎	▃▁·▄▃▂█▂▁▁·▃ beats 4/29	medium · avg 60%	▇ 1.7%	at its level
System One Bench: AG News (English)	86.0% ✎	▁···▅▁▅▂▆█▂▅ beats 4/29	easy · avg 88%	▇ 0.5%	blunt
System One Bench: Banking77 (English)	1.0% ✎	█▅▆█▅▃▃▃···▂ beats 0/26	very easy · avg 87%	—	failed run: All requests failed; not a measurement of ability. Source records unsupported option cardinality.

language understanding▂▂▂▃▂▅▆▃█▆▄▃ #151 of 256 · 13% of its score · 9 tests

Trails Decider 2B, Crivo-1B, vera-core-2 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
System One Bench: OSACT4 offensive (Arabic)	86.0% ✎	▁······▁▃█▇▄ beats 9/29	easy · avg 84%	▇ 3.2%	sharp
System One Bench: BoolQ (English)	87.0% ✎	▁▁▁▃·▁▃▃▇▇█▅ beats 8/29	easy · avg 86%	▇ 3.2%	typical
System One Bench: SST-2 (English)	91.5% ✎	▁·▃▁▂▃▃▅█·▂▁ beats 7/29	easy · avg 92%	▇ 2.1%	typical
System One Bench: SST-5 (English)	53.5% ✎	▃·▁▁▃·▇▄▄▄█▅ beats 11/29	medium · avg 51%	▇ 1.6%	blunt · at its level
System One Bench: XNLI Arabic (Arabic)	58.0% ✎	▄▃▃▁▄▇█▅▃▁·▃ beats 3/29	medium · avg 68%	▇ 0.8%	at its level
System One Bench: XNLI French (French)	67.3% ✎	▂▅·▁▁▅▅█▃▁·▃ beats 5/29	medium · avg 76%	▇ 0.8%	blunt · at its level
System One Bench: Tweet sentiment (Arabic)	71.5% ✎	▁······▁▁▄█▄ beats 21/29	medium · avg 62%	▇ 0.6%	blunt · at its level
System One Bench: XNLI English (English)	74.7% ✎	▃·▃·▃▇█▄█▃▃▁ beats 2/29	easy · avg 85%	▇ 0.2%	blunt · noisy

1 more test

test	result	vs the others in its run	difficulty	counts	why
System One Bench: DAIR Emotion (English)	54.0% ✎	██▃·····▁··▁ beats 2/29	medium · avg 59%	▇ 0.0%	blunt · at its level

knowledge & reasoning▁▂▂▂▂▃▃▅▅▆█▂ #27 of 263 · 36% of its score · 6 tests

Beats Bespoke Nimble v4 Max (27B), vera-large, KnowLine 26B-A4B Gen2.5 here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
LeanMCP · scienceqa text	96.2% ✎	▄··▄█▄█▄·█▄▄ beats 10/11	very easy · avg 88%	▇▇▇▇ 9.3%	sharp
LeanMCP · medqa usmle	83.1% ✎	▄▄▄·▄▄▄▄▄·██ beats 5/11	easy · avg 72%	▇▇▇ 7.7%	typical
MedQA · Freyer et al. (first-valid accuracy)	82.6% ✎	█··▃▃·▃▃▃·▃█ beats 5/11	easy · avg 72%	▇▇ 6.0%	typical
LeanMCP · medmcqa	72.2% ✎	▃·▃··▃█▃·▅▅▃ beats 3/11	medium · avg 63%	▇▇ 4.9%	at its level
LeanMCP · pubmedqa	59.4% ✎	▂···▃·▂·▄▄▂█ beats 1/11	medium · avg 54%	▇▇ 4.9%	at its level
LeanMCP · mmlu pro	72.5% ✎	▄··▄▄████··▄ beats 10/11	medium · avg 54%	▇ 3.7%	at its level

workflows▁▁▁▂▁▂▂▆█▄▁▁ #36 of 114 · 30% of its score · 10 tests

Beats vLLM-SR d3-flash (9B), Clef-omni, Jebadiah 27B v2.1 (Ollaya Q4_K_M) here, though they rank higher overall. Trails StartLux-Decision v1.1 2B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LeanMCP · rewardbench2	75.0% ✎	▄▄▄·▄▄▄▄█·█▄ beats 8/11	medium · avg 58%	▇▇ 4.3%	at its level
LeanMCP · mcpmark tool selection	94.4% ✎	▃··········█ beats 1/11	very easy · avg 96%	▇▇ 4.2%	sharp · maxed out · noisy
LeanMCP · mcpmark server routing	47.9% ✎	▂··▂··▄█▂·▂▄ beats 5/11	medium · avg 36%	▇▇ 4.2%	at its level
System One Bench: Typed decisions (English)	69.9% ✎	▁▁·▁▁▁▅▄▇██▁ beats 13/29	medium · avg 63%	▇▇ 4.0%	blunt · at its level
LeanMCP · aegis2 response	75.4% ✎	█··▃·▃·█··▅▅ beats 3/11	easy · avg 70%	▇ 3.7%	at its level
LeanMCP · atbench500	73.8% ✎	▃··▅▅█··▅··▅ beats 3/11	medium · avg 63%	▇ 2.8%	sharp · at its level · noisy
LeanMCP · helpsteer2 helpfulness	1.0720 ordinal levels ✎	▄▄▄▄▄·▄·█▄▄▄ beats 3/10	avg 1.357 ordinal levels	▇ 2.5%	blunt
LeanMCP · sst5	0.5990 ordinal levels ✎	▂·····▃▂▄█▄▂ beats 1/11	avg 0.879 ordinal levels	▇ 2.4%	blunt

2 more tests

test	result	vs the others in its run	difficulty	counts	why
LeanMCP · prompt injections	70.7% ✎	▂▂▂▃█·▄▂···▂ beats 3/11	medium · avg 62%	▇ 1.9%	at its level
LeanMCP · jailbreak classification	96.5% ✎	▂·▄··▂···█▆▂ beats 4/11	very easy · avg 95%	▇ 0.1%	blunt · maxed out · noisy

$ where to run it

Through an API

api · OpenRouter · price unknown · openrouter.ai ↗

Model repository and documentation: openrouter.ai ↗

$ how the score was worked out · LeanMCP · JevBench 62%, System One Bench (Ali Alhashi… 31%, Freyer et al. · MedQA 6%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
LeanMCP · JevBench	16	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 62%	MIT
System One Bench (Ali Alhashimi)	15	▇▇▇▇▇▇▇▇▇ 31%	CC-BY-4.0
Freyer et al. · MedQA	1	▇▇ 6%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 56.5

not counted (1)

System One Bench (Ali Alhashimi)

System One Bench: Banking77 (English) [correct]

1.0%

failed run: All requests failed; not a measurement of ability. Source records unsupported option cardinality.

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-11-03 02:37 UTC · sources and licences · json