Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →There is no single winner across this benchmark table. The ai-sage Hugging Face model card reports GigaChat 3.5 Ultra Reasoning slightly ahead on two AIME rows, while DeepSeek V4 Flash Preview Reasoning leads several other listed tasks and has the higher reported average. To interpret the comparison, compare models within the same benchmark row, then check how that row was evaluated; scores from different benchmarks are not interchangeable.
How to read a benchmark table
A benchmark row measures performance on a particular task under a particular evaluation setup. Read the task name, score, aggregation rule and any notes together. A score of 90 on one benchmark does not automatically mean the model is more capable than a score of 80 on another: the tasks, scales and procedures may differ.
As an Amazon Associate I earn from qualifying purchases.
- Identify the task and score direction. Check what the benchmark measures and whether higher scores indicate better results.
- Compare within a row. Treat each model’s result as comparable only for that benchmark and its stated setup.
- Check the aggregation and evaluation notes. Sample counts, prompting, judges, tools and time limits can affect the result.
- Keep missing values missing. A dash is not a zero and should not be included as though it were a measured score.
- Do not infer significance from a small gap alone. Without uncertainty intervals or enough repeated-run detail, a small difference may not be robust.
Which model scored higher in this comparison?
The figures below are reported by the ai-sage Hugging Face model card; they were not independently reproduced here. The card does not establish a common real-world meaning for score differences across benchmarks.
Free tools Windows power users keep installed
One-click scans. No signup required.
STEM benchmarks
| Benchmark | GigaChat 3.5 Ultra Reasoning | DeepSeek V4 Flash Preview Reasoning | Higher reported score |
|---|---|---|---|
| AIME 2025 (mean@32) | 89 | 88.95 | GigaChat |
| AIME 2026 (mean@32) | 92 | 90.4 | GigaChat |
| HMMT 2025 (mean@8) | 83.13 | 95.21 | DeepSeek |
| IMOAnswerBench | 73 | 85.75 | DeepSeek |
| GPQA-Diamond | 82.32 | 87.4 | DeepSeek |
General benchmarks
| Benchmark | GigaChat 3.5 Ultra Reasoning | DeepSeek V4 Flash Preview Reasoning | Higher reported score |
|---|---|---|---|
| IFBench | 77 | 73.33 | GigaChat |
| StructEval | 85 | 80.19 | GigaChat |
| MERA-2.0 | 42.3 | not stated (ai-sage Hugging Face model card) | Not comparable from the listed values |
| Function Calling V4 | 58.59 | 68.06 | DeepSeek |
| TAU3-bench | 47.8 | 67.7 | DeepSeek |
| Natural Plan | 80.19 | 88 | DeepSeek |
Code benchmarks
| Benchmark | GigaChat 3.5 Ultra Reasoning | DeepSeek V4 Flash Preview Reasoning | Higher reported score |
|---|---|---|---|
| Live Code Bench v6 | 85.4 | 87.87 | DeepSeek |
| SWE-bench Verified | 64.7 | 78.6 | DeepSeek |
| Terminal-Bench 2 | 30.3 | 56.6 | DeepSeek |
Arena results
| Benchmark | GigaChat 3.5 Ultra Reasoning | DeepSeek V4 Flash Preview Reasoning | Higher reported score |
|---|---|---|---|
| Arena Hard Logs V3 | 56.5 | 53.7 | GigaChat |
| Arena Hard Ru | 60.7 | 36.8 | GigaChat |
| Ru LLM Arena | 64 | 48.5 | GigaChat |
| Pollux | 49 (Ultra Reasoning); 71.6 (Ultra Instruct) | 67.9 | Ultra Instruct, a different GigaChat variant |
The model card reports averages of 68.88 for GigaChat Ultra Reasoning and 72.71 for DeepSeek. It does not explain cross-task normalization sufficiently to make that average a general-purpose quality score. In particular, the missing DeepSeek MERA-2.0 value should not be treated as zero or silently folded into a comparison.
#1 Best Overall
What do mean@32 and mean@8 mean?
These labels indicate that the reported result is an average across multiple generated samples: mean@32 refers to 32 samples, and mean@8 to eight. The model card labels AIME 2025 and AIME 2026 as mean@32, and HMMT 2025 as mean@8. The aggregation method matters because a result based on multiple attempts is not the same kind of measurement as a single answer. Do not assume every benchmark in the table uses the same sample count or aggregation rule.
Which evaluation details affect the comparison?
The model card lists several protocol details that help qualify its results:
- IMOAnswerBench uses Qwen-3-235B-Instruct-2507 as judge.
- TAU3-bench averages Airline, Retail, Telecom and Banking.
- Natural Plan uses a corrected scorer that normalizes UTF-8 characters to ASCII.
- SWE-bench Verified and Terminal-Bench 2 use mini-swe-agent with a three-hour timeout.
- Arena evaluations use MiniMax-M2.7 as judge and GPT-5.2 as baseline.
- For benchmarks without a methodology-defined system prompt, the model card says an empty system prompt was used.
These details are not incidental: judge choice, prompt, agent setup and time limit help define what the result represents. The separate harness-benchmark project offers a general caution that single runs do not establish repeatability or statistical significance, and that different harnesses or profiles can measure different things. It is not a validation of the scores in this model card.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What does the model card say about reasoning-token use?
The ai-sage Hugging Face model card states that GigaChat 3.5 Reasoning uses 37% fewer reasoning tokens overall than DeepSeek V4 Flash Preview across four reported evaluation samples. Its listed per-benchmark figures are:
| Evaluation | Samples | GigaChat mean tokens | DeepSeek mean tokens | Reported reduction |
|---|---|---|---|---|
| AIME 2025 | 240 | 13,980 | 19,129 | 27% |
| AIME 2026 | 240 | 13,635 | 17,697 | 23% |
| HMMT | 480 | 13,311 | 19,553 | 32% |
| IMOAnswerBench | 1,096 | 17,074 | 29,041 | 41% |
These are source-reported figures, not independently reproduced measurements. The reviewed page does not establish when the measurements were published; the years in the AIME labels identify benchmark versions, not the publication date. Token counts alone also do not establish comparative cost, latency or hardware needs.
What else is known about GigaChat 3.5 Reasoning?
The model card describes GigaChat 3.5 Reasoning as a 432B Mixture-of-Experts model with 28B active parameters and a maximum supported context length of 262K tokens. It provides software inference instructions for frameworks and local tooling. Those specifications may matter when considering practical use, but they do not supply a matched cost, latency or hardware comparison with DeepSeek.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the benchmark table can and cannot tell you
The table supports a task-by-task comparison, not a universal ranking. GigaChat edges the two listed AIME results and leads several other rows; DeepSeek leads a number of STEM, general and code rows and has the higher reported average. The model card does not provide enough information to establish that its small score gaps are statistically meaningful, that every evaluation used fully comparable conditions, or that these results represent every release or service configuration.
For provenance, the scores come from an ai-sage Hugging Face repository. The available material does not confirm that account as an official publisher for either model developer, and it does not independently verify the results or give complete protocols for every benchmark. Keep the competitor name exact: the table compares GigaChat 3.5 Ultra Reasoning with DeepSeek V4 Flash Preview Reasoning, not every DeepSeek version.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
