GigaChat3.5-Reasoning

GigaChat 3.5 Reasoning is the first GigaChat model with full reasoning trained with online RL. Compared with GigaChat 3.5 Ultra Instruct, the largest gains are in mathematics, code, instruction following, and structured output.

This repository contains GGUF weights for llama.cpp.

Version for high-performance inference in FP8 - GigaChat3.5-432B-A28B-Reasoning.

Model in BF16 - GigaChat3.5-432B-A28B-Reasoning-bf16.

Model architecture

GigaChat 3.5 Reasoning is a 432B Mixture-of-Experts model with 28B active parameters. It uses a custom hybrid architecture that combines Multi-head Latent Attention (MLA) with GatedDeltaNet linear-attention layers.

The model also uses GatedNorm, a learned multiplicative gate applied after RMSNorm, and has three MTP heads for speculative decoding. The maximum supported context length is 262K tokens.

Online RL

Post-training starts from an SFT checkpoint. We train six domain experts independently with online RL and then combine them into one release model with on-policy distillation (OPD).

The experts are trained with CISPO. Before training, the current checkpoint is evaluated on the task pool and tasks solved in more than 75% of attempts are removed. As the model improves, the training set shifts toward harder tasks.

Rewards are domain-specific but follow the same general construction: gated checks for hard constraints, additive rewards for answer quality, and an adaptive length penalty.

After RL, the six experts are combined with on-policy distillation. The student generates its own trajectory, while the expert for the corresponding domain provides token-level supervision on that trajectory.

Benchmark scores

* IMOAnswerBench uses Qwen-3-235B-Instruct-2507 as the judge. ** TAU3-bench is averaged across 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