whatmodel.sh / U2 Flash

U2 Flash is a closed-weights language model from Unisound, released September 2026. Licence: proprietary.

Catalogue identity notes

Unisound (云知声) U2-Flash, API model id 'u2-flash', released on Unisound's open platform 2026-09-15 (MaaS release log '2026 年 9 月 15 日 … U2-Flash 发布'); text in/out, sparse MoE, thinking strength none/low/high/max. Generally available via Unisound MaaS API and Token Plan. Corresponds to Arena webdev entry 'U2 Flash' (org 'unisound', Proprietary). Licence: No weights found: Unisound's release log calls U2-Decision 'open-sourced' (开源) but makes no such statement for U2-Flash; no Hugging Face repo matches 'U2-Flash' (HF API search, 2026-11-12). Arena lists it Proprietary. Context window not recorded (Unisound MaaS model page key-info table: '上下文窗口 | 1M', '最大输出 | 128K'. Exact token count not stated; 1M read as 1,000K. No models.dev toml at 5056efa. Observed 2026-11-12.). Price: No USD price published. Unisound MaaS price page (https://maas.unisound.com/docs/price/overview) lists 'U2-Flash | 0.6 | 0.06 | 1.2' CNY per 1M tokens (input / cache hit / output), observed 2026-11-12; a separate 'recharge 40% off' (限时6折) promotion is shown. Numeric USD fields left empty; no models.dev toml.

Model information sources

Release date ↗

Not recorded: Unisound's release log says 'total parameters about 266B, about 10B activated per inference' (总参数约 266B、单次推理仅激活约 10B) while the product page says '300B规模' (300B scale). Approximate and inconsistent, so params_b/active_b omitted.

75.8±2.3#29 of 974 LLMsfrom 1 benchmark result across 1 source* thinly tested, treat the rank as provisional

Score within the LMArena (text) comparison set only.

$ how it compares

The LLMs ranked either side of U2 Flash, out of 1206 with a score.

#modelscorewhereRAMreleased
27Qwen3.8 Max (2.4T-A95B) Alibaba76±1api $3.00/M +local~594GB2026-08
28Claude Opus 4.8 Anthropic76±0api $10.00/M2026-05
29U2 Flash Unisound76±2api price unknown2026-09
30GPT-5.2 Pro OpenAI76±1api $57.75/M2025-12
31Claude Haiku 5.5 Anthropic76±1api $0.20/M2026-10

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasksnot ranked on this task yet · 3 tests
testresultvs the others in its rundifficultycountswhy
SWE-bench Pro (split unspecified) lab61.6% ✎▁·▁▁▁▁▁▄█▁▁▁ beats 49/64very hard · avg 4.0%—not counted yet: no other model in the same run to compare it with
DeepSWE v1.1 (resolved) lab64.6% ✎▂····▃▁▁▄█▆█ beats 13/36—not counted yet: no other model in the same run to compare it with
Terminal-Bench 3.0 lab24.3% ✎█·█··██···██ beats 4/5—not counted yet: no other model in the same run to compare it with
coding▁▁▁▃▆▇█▆▆▃▃▁ #33 of 487 * · 98% of its score · 3 tests

Beats GPT-5.5, GPT-5.6 Terra, Claude Opus 4.8 here, though they rank higher overall. Trails Qwen3.8-Flash-Next, Grok 4.7, Qwen3.8 Max (0902) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1612.8 ✎▁▂▄▃▇█▅▇▇▅▂▁ beats 103/134avg 996▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 98%sharp
SWE-bench Pro (split unspecified) lab61.6% ✎▁·▁▁▁▁▁▄█▁▁▁ beats 49/64very hard · avg 4.0%—not counted yet: no other model in the same run to compare it with
DeepSWE v1.1 (resolved) lab64.6% ✎▂····▃▁▁▄█▆█ beats 13/36—not counted yet: no other model in the same run to compare it with
mathnot ranked on this task yet · 2 tests
testresultvs the others in its rundifficultycountswhy
IMO-AnswerBench lab90.3% ✎▁▁·▁▁▁▄▆█▄▃▃ beats 41/45medium · avg 42%—not counted yet: no other model in the same run to compare it with
HMMT February 2026 lab97.1% ✎▃▃▁▄▇▇▇█▅▃▃▃ beats 38/38hard · avg 31%—not counted yet: no other model in the same run to compare it with
other tests1 not in a task · 0.0% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond lab92.4% ✎▁▁▁▁▂▃▅▆▇▇█▂ beats 306/336hard · avg 39%—not counted yet: no other model in the same run to compare it with

$ where to run it

Through an API

api · Unisound MaaS · price unknown · maas.unisound.com ↗

$ how the score was worked out · LMArena leaderboard dataset 98%

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

sourceresultsshare of the scorelicence
LMArena leaderboard dataset1CC BY 4.0
prior · pull toward the middle
weighted average= 75.8

not counted (6)

Unisound MaaS model page (lab-reported scores)GPQA Diamond [correct]92.4%fewer than three models have this test yet, or this source ran it for this model alone
Unisound MaaS model page (lab-reported scores)HMMT February 2026 [correct]97.1%fewer than three models have this test yet, or this source ran it for this model alone
Unisound MaaS model page (lab-reported scores)IMO-AnswerBench [correct]90.3%fewer than three models have this test yet, or this source ran it for this model alone
Unisound MaaS model page (lab-reported scores)DeepSWE v1.1 (resolved) [correct]64.6%fewer than three models have this test yet, or this source ran it for this model alone
Unisound MaaS model page (lab-reported scores)SWE-bench Pro (split unspecified) [correct]61.6%fewer than three models have this test yet, or this source ran it for this model alone
Unisound MaaS model page (lab-reported scores)Terminal-Bench 3.0 [correct]24.3%fewer than three models have this test yet, or this source ran it for this model alone

$ improve this page

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 · sources and licences · json