whatmodel.sh / GigaChat 3.5 Ultra Reasoning (432B-A28B)

GigaChat 3.5 Ultra Reasoning (432B-A28B)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

GigaChat 3.5 Ultra Reasoning (432B-A28B) is an open-weights language model from Sber, released September 2026. Its published size is 438.09B parameters (28B active per token) and a 262k-token context window. Licence: mit.

Catalogue identity notes

ai-sage/GigaChat3.5-432B-A28B-Reasoning @02f5537 (created 2026-09-03, licence mit). Sber's first GigaChat with full reasoning trained with online RL (Sept 2026), FP8 weights; 432B MoE with 28B active. Lab rows from the GigaChat 3.5 Reasoning card table (with Sber's run of the catalogued DeepSeek V4 Flash Preview).

Model information sources

Release date ↗ · Weights licence ↗

HF safetensors metadata 438,085,063,424 parameters; 28B active per the maker card.

66.7±1.4#119 of 719 LLMsfrom 10 benchmark results across 1 source

$ how it compares

The LLMs ranked either side of GigaChat 3.5 Ultra Reasoning (432B-A28B), out of 828 with a score, then more from Sber.

#modelscorewhereRAMreleased
136Gemini 2.5 Pro Preview (03-25) Google67±2comparison only2025-04
137Gemini 2.5 Pro Preview (05-06) Google67±1comparison only2025-05
138GigaChat 3.5 Ultra Reasoning (432B-A28B) Sber67±1local~461GB2026-09
139Qwen3.5 35B-A3B Alibaba67±0api $0.35/M +local~23GB2026-02
140DeepSeek V3.1 Terminus DeepSeek67±1local~230GB2025-09
more from Sber
321GigaChat 3.5 Ultra Instruct (432B-A28B) Sber55±1local~457GB2026-07

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #125 of 300 · 27% of its score · 3 tests

Beats o3, Gemini 2.5 Pro, Gemini 3.1 Flash-Lite here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Large, Nova 2 Pro (Preview) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
SWE-bench Verified lab64.7% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 64/131hard · avg 24%▇▇▇▇▇ 13%sharp · at its level · small run
Terminal-Bench 2.0 lab30.3% ✎▂▂▅▅████▆▂▃▃ beats 5/36very hard · avg 9.5%▇▇▇▇ 11%at its level · small run
τ³-bench lab47.8% ✎▁▁▂▁·▁▁·▃▁▂█ beats 9/25hard · avg 34%▇ 3.1%blunt · at its level · small run
coding▁▁▁▃▆██▇▆▄▃▁ #114 of 458 * · 22% of its score · 2 tests

Beats Qwen3.5 Plus, Gemini 2.5 Pro, Kimi K2 Thinking here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Solar Pro 4, Laguna XS 2.1 (33B-A3B) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
SWE-bench Verified lab64.7% ✎▁▁▁▁▁▁▂▄█▆▁▁ beats 64/131hard · avg 24%▇▇▇▇▇ 13%sharp · at its level · small run
LiveCodeBench v6 (window unspecified) lab85.4% ✎▂▃▄▄▅██▅▄▇▃▆ beats 52/65medium · avg 46%▇▇▇▇ 9.4%small run
math▁▂▃▅▆▅▄█▄▁▁▁ #88 of 398 · 42% of its score · 4 tests

Beats Claude Sonnet 4.6, Claude Opus 4.5, MiniMax-M3 here, though they rank higher overall. Trails Gemini 2.5 Flash, Granite 4.2 30B, INTELLECT-3 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
HMMT 2025 (edition unspecified) lab83.1% ✎▃█▇▅▇▅▁·▁··▁ beats 26/32hard · avg 30%▇▇▇▇ 11%sharp · small run
AIME 2025 lab89.0% ✎▁·▁▁▁▂▅▆█▇▅▁ beats 110/133medium · avg 48%▇▇▇▇ 11%sharp · small run
AIME 2026 lab92.0% ✎▁▁▁▂▁▂▆█▆▂·▁ beats 27/49medium · avg 53%▇▇▇▇ 10%sharp · small run
IMO-AnswerBench lab73.0% ✎▁▁·▁▁▁▄▅█▄▂▃ beats 11/42hard · avg 31%▇▇▇▇ 9.4%at its level · small run
other tests2 not in a task · 21% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond lab82.3% ✎▁▁▁▁▂▄▅▅███▄ beats 212/310hard · avg 45%▇▇▇▇▇▇ 16%sharp · small run
IFBench (aggregation unspecified) lab77.0% ✎▃▇▂▄▅▅▂▄▄▆█▇ beats 69/80medium · avg 40%▇▇ 4.8%at its level · small run

$ where to run it

Other ways to run it

  1. Run it as published · needs ~438 GB, NVIDIA only · from huggingface.co/ai-sage
    vllm serve ai-sage/GigaChat3.5-432B-A28B-Reasoning --tensor-parallel-size 8

Weights and documentation: huggingface.co/ai-sage ↗

$ how the score was worked out · ai-sage/GigaChat3.5-432B-A28B… 99%

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
ai-sage/GigaChat3.5-432B-A28B-Reasoning model card · lab-reported10▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 99%MIT
prior · pull toward the middle▇ 1%
weighted average= 66.7

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit 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 · sources and licences · json