Sber's open models: a 432B chat model, speech and embeddings

Sber's GigaChat team has five MIT-licensed entries on Hugging Face: GigaChat 3.5 Ultra, 3.1 Lightning, GigaAM-v3, GigaAM Multilingual and Giga-Embeddings. Sizes, scores and caveats.

  • Analysis
  • 3 min read
  • Published

The GigaChat team at Sber publishes open weights under the ai-sage name on Hugging Face. Between November 2025 and September 2026 it released models that cover chat, speech recognition and embeddings for Russian, and we list five entries, each marked MIT. This piece sets them side by side, and every score in it is Sber’s own.

The five entries

Model Job Size Announced or first on Hugging Face Smallest build we list
GigaChat 3.5 Ultra Chat, reasoning, agents 432B mixture of experts, 28B active 6 July 2026 (Reasoning 3 September) FP8, 437.9 GB
GigaChat 3.1 Lightning Small chat model 10B mixture of experts, 1.8B active 24 March 2026 GGUF Q4_K_M, 6.5 GB
GigaAM-v3 Russian speech recognition 220M to 240M 19 November 2025 About 0.45 GB per checkpoint
GigaAM Multilingual Russian and Central Asian speech recognition 220M, 600M 14 July 2026 0.88 GB
Giga-Embeddings Russian and English embeddings 480M, 3B, 10B-A1.8B 29 July to 10 August 2026 0.98 GB

The chat models

GigaChat 3.5 Ultra replaces the 700B GigaChat 3.1 Ultra with a model that the card says is about 40% smaller and stronger in code, mathematics and agent tasks. A custom hybrid attention 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. It comes as Instruct and Reasoning versions, with 262,144 positions in the configuration. The FP8 files are 437.9 GB and 446.0 GB, and the cards tell you to build SGLang or vLLM from specific pull requests, so it is a project for a multi-GPU server.

In Sber’s table the Reasoning version is marginally ahead of DeepSeek V4 Flash Preview Reasoning on AIME 2025 and 2026, with 89 and 92 against 88.95 and 90.4. It trails 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.

GigaChat 3.1 Lightning is the small one. Only 1.8B of its 10B parameters are active, but all 10B must sit in memory, and the files are 12.3 GB in FP8, 23.0 GB in BF16 and 6.5 GB as a Q4_K_M GGUF. Sber’s table gives it an average of 0.631 over seven tests against 0.612 for Qwen3-4B-Instruct, with the lead on Russian tests such as MMLU RU (0.680 against 0.597). Qwen3-4B-Instruct is ahead on MMLU-Pro, BBH and HumanEval Plus.

Speech and embeddings

GigaAM-v3 averages 8.4% word error rate for its RNN-T model and 9.2% for its CTC model over five Russian test groups, against 25.1% for Whisper. The gap is widest on disordered speech (19.2% against 59.3%) and call-center audio (9.5% against 23.9%), and the two e2e checkpoints write punctuated, normalized text. GigaAM Multilingual adds Kazakh, Kyrgyz and Uzbek. On FLEURS its large model scores 4.4%, 5.5% and 7.3% on them, where Whisper large v3 scores 32.4%, 86.3% and 105.4%. On English it trails, at 9.4% against 3.9%.

Giga-Embeddings comes in 480M, 3B and 10B-A1.8B sizes. The 480M model scores 70.98 on MTEB Russian against 63.64 for Qwen3-Embedding-0.6B, which leads it on English, code and multilingual tests. The 3B and 10B models score 74.57 and 74.99 on Russian.

What to check before you use them

  • Licenses. The cards say MIT, but the two GigaChat repositories contain no LICENSE file, so the license rests on the card metadata and Sber’s Habr announcement. Keep both with your records.
  • Hardware. Ultra needs a server. Everything else is under 25 GB per file, and Lightning also has the 12.3 GB and 6.5 GB builds. The VRAM calculator gives a first estimate.
  • Setup. GigaAM and Giga-Embeddings need trust_remote_code. The GigaAM-v3 card pins Transformers 4.57.1 and PyTorch 2.8, and GigaAM handles audio of up to 25 seconds before its long-form mode, which uses pyannote.audio, takes over. Giga-Embeddings queries need an “Instruct: … Query: …” prefix and mean pooling.
  • Evidence. The test sets differ, GigaAM Multilingual’s exclude references that contain digits, and Gemini 2.5 Pro judged the e2e comparison with Whisper.

See the lists of speech recognition models and embedding models, and test on your own Russian data using how to evaluate models yourself.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
GigaChat 3.5 Ultra Sber (GigaChat) 434B – 438B 28B active 256K MIT (Permissive) 299 GB Big-memory desktop
GigaChat 3.1 Lightning Sber (GigaChat) 11.5B 1.8B active 256K MIT (Permissive) 7.15 GB Laptop or phone
GigaAM-v3 Sber (GigaChat) 220M – 240M – MIT (Permissive) 0.506 GB Laptop or phone
GigaAM Multilingual Sber (GigaChat) 220M – 600M – MIT (Permissive) 1.01 GB Laptop or phone
Giga-Embeddings Sber (GigaChat) 480M – 10.5B 1.8B active 8K MIT (Permissive) 1.13 GB Laptop or phone