GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
dkagramanyan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
dkagramanyan
Model Tree
ai-sage/GigaChat3.1-10B-A1.8B-bf16
Input Modalities
Output Modalities
Supported Functionality
lm_head, MTP layer.512 samples, 897k tokens, max length 2048, chat template applied.
| Dataset | Lang | Samples |
|---|---|---|
| Vikhrmodels/GrandMaster-PRO-MAX | RU | 192 |
| Vikhrmodels/Grounded-RAG-RU-v2 | RU | 96 |
| HuggingFaceH4/ultrachat_200k | EN | 128 |
| NousResearch/hermes-function-calling-v1 | EN | 64 |
| ise-uiuc/Magicoder-Evol-Instruct-110K | EN | 32 |
lm-evaluation-harness, vLLM 0.25, RTX 3090, greedy, 500 samples per task.
| Task | bf16 | W4A16 | Δ |
|---|---|---|---|
| gsm8k (5-shot) | 85.40 | 82.20 | -3.20 |
| ifeval (prompt strict) | 71.40 | 71.80 | +0.40 |
| mgsm_native_cot_ru (8-shot) | 73.20 | 68.00 | -5.20 |
| mgsm_native_cot_en (8-shot) | 86.40 | 85.60 | -0.80 |
| global_mmlu_full_ru | 64.43 | 63.25 | -1.18 |
| global_mmlu_full_en | 71.80 | 70.57 | -1.23 |
| arc_ru | 50.40 | 50.00 | -0.40 |
| arc_challenge | 60.20 | 59.20 | -1.00 |
| hellaswag_ru | 64.40 | 62.60 | -1.80 |
| hellaswag | 68.60 | 68.60 | 0.00 |
| belebele_rus_Cyrl | 80.80 | 78.00 | -2.80 |
| belebele_eng_Latn | 88.40 | 87.20 | -1.20 |
| Mean | -1.53 |
bash
docker run --rm --gpus all --ipc=host -p 8000:8000 -v ~/.cache/huggingface:/root/.cache/huggingface \ -e VLLM_USE_DEEP_GEMM=0 vllm/vllm-openai:v0.25.0 your-hf-user/GigaChat3.1-10B-A1.8B-W4A16 \ --max-model-len 16384 --enable-auto-tool-choice --tool-call-parser gigachat3License: MIT, as the base model.
docker run --rm --gpus all --ipc=host -p 8000:8000 -v ~/.cache/huggingface:/root/.cache/huggingface \ -e VLLM_USE_DEEP_GEMM=0 vllm/vllm-openai:v0.25.0 your-hf-user/GigaChat3.1-10B-A1.8B-W4A16 \ --max-model-len 16384 --enable-auto-tool-choice --tool-call-parser gigachat3