Skip to content
Other

SGI-Preview

70B dense-модель для агентного кодинга

2 min read

SGI-Preview — instruct-модель на 70B параметров от SGI Team. Это dense Transformer, обученный с нуля на 12 трлн токенов, с заявленной поддержкой контекста до 1M. В карточке сделан акцент на агентном кодинге и reasoning.

Ключевые особенности

  • Dense-архитектура: 70B параметров, 80 слоёв, узкий d_model 5120; авторы сознательно выбрали её вместо MoE.
  • Terminal-Bench 2.1 — 80.5; по этому бенчмарку модель выше Reflection Beam 501B (80.1), Muse Spark Open 816B (76.4), Inkling 975B (63.8) и Nemotron 3 Ultra 550B (56.4).
  • Заявлен контекст до 1M; обучение включало midtraining, расширение контекста, SFT, long-context SFT и RL.
  • Есть BF16- и NVFP4-версии; NVFP4-чекпоинт позиционируется для рабочих станций вроде NVIDIA DGX Spark.
  • Лицензия Apache 2.0; отдельно выпущена pretrain-only базовая модель SGI-Base-Preview.

Пример

Сценарий, под который заточены бенчмарки, — агент в терминале: модель чинит код и проходит тесты в задачах уровня Terminal-Bench. Например, запрос: Найди причину падения тестов в репозитории и предложи патч. Модель рассчитана на многошаговое reasoning и агентные задачи.

ПодробнееTerminal-Bench 2.1: 80.5 (Terminus 2 — 80.5), SWE-bench Verified — 76.5, MMLU — 89.6.
70B dense; NVFP4-чекпоинт рассчитан на рабочую станцию, например NVIDIA DGX Spark; ~240 tok/s на 2×B200 при batch=8.
Претрейн — 12T токенов с преобладанием английского; заявленный контекст — до 1M; лицензия Apache 2.0.
Discuss Telegram 0 29

Will it run?

Only with RAM offload

Even 4-bit is larger than your VRAM. Some layers can stay in system RAM (llama.cpp, ComfyUI), but it will be several times slower.

  • BF16weights 121 GB · ready files133 GB
  • FP8 / INT8recommendedweights 60 GB67 GB
  • 4-bitweights 34 GB38 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # xsingh/SGI-Preview
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download xsingh/SGI-Preview --local-dir ~/models/SGI-Preview
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to SGI-Preview
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of SGI-Preview (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

xsingh/SGI-Preview
Author
xsingh
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

290

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments