ai-sage/GigaChat3.5-432B-A28B-GGUF

GigaChat 3.5 Ultra

GigaChat 3.5 Ultra is the flagship instant model of the GigaChat family. It is a large-scale Mixture-of-Experts (MoE) model with 432B total parameters, built on a custom hybrid attention architecture that combines Multi-head Latent Attention (MLA) with GatedDeltaNet linear-attention layers. The model targets multilingual assistant workloads, reasoning, code, agentic/tool-use scenarios, and large-cluster deployment.

Compared to the previous flagship GigaChat 3.1 Ultra (700B), version 3.5 is ~40% more compact yet stronger in code, mathematics, and agentic scenarios. It also uses roughly 4× less KV-cache per token, fits more than 2× more context into the same memory, and improves generation throughput by ~20%.

Version for high-performance inference in fp8 - GigaChat3.5-432B-A28B.

Model in bf16 is GigaChat3.5-432B-A28B-bf16.

Base version for training - GigaChat3.5-432B-A28B-base.

Training checkpoints - GigaChat3.5-432B-A28B-checkpoints.

More details can be found in the Habr article.

Model architecture

GigaChat 3.5 Ultra uses a custom MoE architecture. The core change relative to 3.1 is a self-designed hybrid architecture and a matching training recipe: every acceleration feature (linear attention, MTP) was paired with a stabilizing mechanism so the model could be trained to full scale without loss of stability.

Mixture-of-Experts (MoE)

The model has 432B total and 28B active parameters, keeping inference cost far below that of an equally large dense model. The MoE decoder layer is composed of attention, the MoE (expert) block, and a post-normalization applied before the residual add.

Hybrid attention: MLA + GatedDeltaNet

Standard attention grows more expensive with context length: the longer the request, the larger the KV-cache, and the more generation is bottlenecked on memory. GigaChat 3.5 introduces a hybrid design in which some layers remain regular MLA and the rest are linear-attention layers based on GatedDeltaNet. This preserves the strengths of full attention while lowering the cost of long context.

Gated Normalization (GatedNorm)

Large models tend to develop implicit self-stabilization (attention/residual sinks), routing most of the signal through a single token or feature to hold the activation scale — which is poorly controlled and can itself become a source of noise at scale. GatedNorm replaces these implicit anchors with an explicit multiplicative gate after RMSNorm, letting the network rescale the signal across features directly. It is made scale-neutral at init via the 2 · sigmoid reparametrization (a plain sigmoid starts near 0.5 and would halve the scale; the factor 2 keeps the gate near 1.0), so it barely perturbs the data flow at start and learns where to attenuate.

Multi-Token Prediction (MTP)

GigaChat Ultra 3.0 had a single MTP head; in GigaChat Ultra 3.5 we added two MTP heads. Greedy decoding accelerates the generation speed ~1.5× with one head and up to 2.2× with two.

Precision and optimizer

The model was trained in native FP8 across all training stages. We also release dequantized bf16 checkpoint.

Alignment

The post-training pipeline runs Stage 1.5 → SFT → DPO → Online RL. Online RL is the headline addition of this release and drove the gains in Instruction Following and on arenas.

Benchmark scores

Base-model

GENERAL

Task	GigaChat 3.1 Base (700B)	GigaChat-3.5-Ultra-Base (430B)	DeepSeek V4 Flash Base (284B)	DeepSeek V3.2 Exp Base (685B)
MMLU (5-shot)	79.89	85.28	88.68	87.47
MMLU-Pro (5-shot)	68.01	74.54	65.86	62.43
GPQA Diamond (official, CoT)	30.3	30.81	22.73	22.22
BBH (3-shot)	83.78	87.5	88.24	89.16
ARC-C (25-shot, acc_norm)	68.34	70.39	72.35	70.31
ARC-E (25-shot, acc_norm)	88.38	88.59	90.82	89.27
HellaSwag (10-shot, acc_norm)	89.43	89.47	88.9	89.28
Winogrande (5-shot)	82.72	85	84.61	84.93
DROP (5-shot, EM)	56.29	59.55	63.88	65.14
TriviaQA (5-shot, EM)	81.4	82.23	83.96	83.88
NQ-Open (5-shot, EM)	37.34	41.66	40.83	42.27
Avg	69.6	72.3	71.9	71.5

MATH

Task	GigaChat 3.1 Base (700B)	GigaChat-3.5-Ultra-Base (430B)	DeepSeek V4 Flash Base (284B)	DeepSeek V3.2 Exp Base (685B)
MATH Minerva (math-verify)	55.78	61.7	54.74	58.2
GSM8K (CoT, math_verify)	86.73	86.58	86.43	84.99
MGSM ru (CoT)	87.6	86	84.4	82
Avg	76.7	78.1	75.2	75.1

CODE

Task	GigaChat 3.1 Base (700B)	GigaChat-3.5-Ultra-Base (430B)	DeepSeek V4 Flash Base (284B)	DeepSeek V3.2 Exp Base (685B)
HumanEval (pass@1)	70.12	80.49	66.46	64.02
HumanEval+ (pass@1)	62.8	75.61	61.59	56.71
MBPP (pass@1)	70.2	70.4	70.2	70.4
MBPP+ (pass@1)	83.33	83.33	77.25	82.28
CRUXEval (pass@1)	64.56	67.5	69.75	69.94
LCB CodeGen Lite	49.29	54.31	57.25	50.24
Avg	66.7	71.9	67.1	65.6

Instruct-model

Task	GigaChat-3.1-Ultra (700B)	GigaChat-3.5-Ultra (430B)	DeepSeek V3.2* (685B)
TAU2-bench	41.87	68.71	66
SWE bench verified **	8.6	42.6	44.8
Terminal bench 2 ***	9	13.48	29.21
Live Code Bench v6	49.29	56.2	59.3
Natural Plan	31.6	27.14	20.44
IFBench	31	43.66	45
MERA Text	71.2	67.3	61.7
Pollux	37.75	65.21	65.6
MMLU-Pro	73.61	75.68	81.1
RubQ_Ru	76.11	81.01	78.33
MATH 500	83	86	91.4
Arena Hard Logs vs GPT-5	55.5	71.4	71.3
Arena Hard Ru vs GPT-5	38.3	69.5	69.3
Ru LLM Arena vs GPT-5	38.2	62.9	63.6
Validator-SBS-Pollux	47.6	70.9	68
Avg	46.2	60.1	61.0

Notes:

• * DeepSeek V3.2 is the instruct version
• ** SWE-bench Verified uses mini-swe-agent, 250 steps
• *** Terminal-Bench 2 uses the terminus-2 agent

Speed

GPU (8 × H100, single stream, 192 tok/prompt)

quant	config	accept rate	decode tok/s	speedup
Q6_K	no MTP	—	47.2	1.00×
Q6_K	MTP-1	0.880	65.4	1.39×
Q6_K	MTP-2	0.775	69.2	1.47×
Q4_K_M	no MTP	—	58.4	1.00×
Q4_K_M	MTP-1	0.858	78.4	1.34×
Q4_K_M	MTP-2	0.759	84.0	1.44×

CPU (Intel Xeon 8462Y+, -t 64, 128 tok/prompt)

quant	config	accept rate	decode tok/s	speedup
Q4_K_M	no MTP	—	7.7	1.00×
Q4_K_M	MTP-1	0.880	8.6	1.12×
Q4_K_M	MTP-2	0.788	9.6	1.25×
Q6_K	no MTP	—	5.7	1.00×
Q6_K	MTP-1	0.876	7.0	1.23×
Q6_K	MTP-2	0.797	7.5	1.32×
Q8_0	no MTP	—	4.8	1.00×
Q8_0	MTP-1	0.886	5.7	1.19×
Q8_0	MTP-2	0.770	6.2	1.29×

Usage Example

Prepare the model

# 1. get the PR
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git fetch origin pull/25342/head:pr-25342
git checkout pr-25342

# 2. download the GGUF (Q8_0 shown; Q4_K_M / Q6_K / bf16 also available)
pip install -U "huggingface_hub[cli]"
hf download ai-sage/GigaChat-3.5-432B-A28B-GGUF \
  --include "GigaChat3.5-432B-A28B-Q8_0/*" \
  --local-dir ./gguf

GPU

Build server

cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j --target llama-server

Start the server

./build/bin/llama-server \
  -m ./gguf/GigaChat3.5-432B-A28B-Q8_0/GigaChat3.5-432B-A28B-Q8_0-00001-of-00010.gguf \
  -ngl 99 \
  -fa on \
  -c 32768 \
  -np 4 \
  -ctk q8_0 -ctv q8_0 \
  --jinja \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --host 0.0.0.0 --port 8080

CPU

Build server

cd llama.cpp
cmake -B build-cpu -DGGML_CUDA=OFF
cmake --build build-cpu --config Release -j --target llama-server

Start the server

./build-cpu/bin/llama-server \
  -m ./gguf/GigaChat3.5-432B-A28B-Q8_0/GigaChat3.5-432B-A28B-Q8_0-00001-of-00010.gguf \
  -ngl 0 \
  -t $(nproc) \
  -fa on \
  -c 32768 \
  -np 4 \
  -ctk q8_0 -ctv q8_0 \
  --jinja \
  --spec-type draft-mtp --spec-draft-n-max 2 \
  --host 0.0.0.0 --port 8080

Request example

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
  "model": "ai-sage/GigaChat3.5-432B-A28B-Q8_0",
  "chat_template_kwargs": {
      "enable_thinking": false
    },
  "temperature": 0,
  "messages": [
    {
      "role": "user",
      "content": "Какая сейчас погода в Москве?"
    }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Получить информацию о текущей погоде в указанном городе.",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {
              "type": "string",
              "description": "Название города (например, Москва, Казань)."
            }
          },
          "required": ["city"]
        }
      }
    }
  ]
}'