GigaChat 3.5 Ultra
Sber's 432B mixture of experts (28B active) with 262K context, in Instruct and Reasoning versions, under MIT.
In short. GigaChat 3.5 is a very large free AI model from Sber that is strongest in Russian and English. It comes as a version that answers right away and a version that thinks step by step first, and the MIT license allows commercial use. It needs a server with several top-end graphics cards, because even the 8-bit weights are about 440 GB.
Summary
GigaChat 3.5 Ultra is Sber's open-weight flagship. Sber's Habr blog announced it on 6 July 2026 and says the weights are on Hugging Face and GitVerse under MIT. It is a 432B mixture of experts with 28B active parameters. The card says it is about 40% smaller than the 700B GigaChat 3.1 Ultra it replaces, yet stronger in code, mathematics and agent tasks. The attention is a custom hybrid that keeps some layers as multi-head latent attention and turns the rest into GatedDeltaNet linear attention, which Sber says cuts the cache per token about fourfold, fits more than twice the context into the same memory and raises throughput by about 20%. The configuration shows 40 layers, 256 routed experts with 8 used per token plus one shared, and 262,144 positions, and the Reasoning card states a 262K maximum context. The model was trained natively in FP8, and multi-token-prediction heads, two on the Instruct card and three on the Reasoning card, speed up decoding.
There are two post-trained versions. Instruct answers directly. GigaChat 3.5 Reasoning, which first appeared on Hugging Face on 3 September 2026, thinks first, and Sber calls it the first GigaChat trained with online reinforcement learning, using six domain experts merged by on-policy distillation. In Sber's own tables, Instruct scores 68.71 on TAU2-bench and 42.6 on SWE-bench Verified against 66 and 44.8 for DeepSeek V3.2, with averages of 60.1 and 61.0. Reasoning scores 89 and 92 on AIME 2025 and 2026, level with DeepSeek V4 Flash Preview Reasoning at 88.95 and 90.4, but it trails that model on GPQA Diamond (82.32 against 87.4), SWE-bench Verified (64.7 against 78.6) and Terminal-Bench 2 (30.3 against 56.6), and its average is 68.88 against 72.71. Sber says Reasoning uses 37% fewer reasoning tokens than that model on four math sets. Several tests are Russian-language, including MERA, RuBQ and the Ru LLM Arena, and all of the numbers are the developer's own.
The FP8 weights are 437.9 GB for Instruct and 446.0 GB for Reasoning, and the BF16 version of Instruct is about 868 GB, so the cards' examples serve it across eight GPUs with expert parallelism, and the Reasoning card names eight H100s. The cards tell you to build SGLang or vLLM from specific pull requests, and Sber also publishes GGUF versions, the base model and training checkpoints. The repositories contain no LICENSE file, so the MIT license rests on the card metadata and Sber's announcement. In the 30 days to 11 November 2026, Hugging Face counted only about 640 downloads for Instruct and 750 for Reasoning, which is typical for a model this size.
Good for
- Russian-language assistants and analysis on your own servers, with a permissive MIT license.
- Agent workflows with function calling, where Sber ships tool-call parsers for SGLang and vLLM.
- Long-context serving on a cluster, because the hybrid attention shrinks the cache per token.
- Research and fine-tuning, since the base model and training checkpoints are published.
Watch out for
- The 8-bit weights alone are about 440 GB, so it needs a multi-GPU server, and the serving instructions depend on unmerged pull requests.
- All benchmarks are Sber's own, and the DeepSeek models it is compared with lead on several, including GPQA Diamond, SWE-bench Verified and Terminal-Bench 2.
- The tagged languages are Russian and English, and the cards do not document the quality of other languages.
- The repositories have no LICENSE file, so keep Sber's announcement and the card metadata with your records.
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Instruct FP8 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Instruct FP8 FP8 | 434B 28B active | 998 GB | 499 GB | 299 GB | Big-memory desktop |
| Instruct BF16 BF16 | 434B 28B active | 998 GB | 499 GB | 299 GB | Big-memory desktop |
| Reasoning FP8 FP8 | 438B 28B active | 1.01 TB | 504 GB | 302 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is GigaChat 3.5 Ultra?
- GigaChat 3.5 is a very large free AI model from Sber that is strongest in Russian and English. It comes as a version that answers right away and a version that thinks step by step first, and the MIT license allows commercial use. It needs a server with several top-end graphics cards, because even the 8-bit weights are about 440 GB.
- Can I use GigaChat 3.5 Ultra commercially?
- Yes. GigaChat 3.5 Ultra is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is GigaChat 3.5 Ultra open source?
- GigaChat 3.5 Ultra is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run GigaChat 3.5 Ultra on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Instruct FP8, needs about 299 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was GigaChat 3.5 Ultra released?
- GigaChat 3.5 Ultra was released by Sber (GigaChat) on July 6, 2026.
- How long a context window does GigaChat 3.5 Ultra have?
- GigaChat 3.5 Ultra supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- GigaChat 3.5 Ultra model card (Hugging Face) huggingface.co
- GigaChat 3.5 Reasoning model card (Hugging Face) huggingface.co
- GigaChat 3.5: smaller, faster, stronger (Sber, Habr, 6 July 2026) habr.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server | |
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| MiniMax M3 MiniMax | 428B 23B active | 1M | MiniMax Community (Conditional) | 295 GB Big-memory desktop | |
| GLM-5.3 Z.ai (Zhipu AI) | 753B 40B active | 1M | GLM-5.3 License (Conditional) | 520 GB GPU server |