Skip to content
Computer vision

B6-27B

LLM, которая сначала ищет, потом отвечает

2 min readupdated

B6-27B от Schneewolf Labs — это B5-27B на 27,78 млрд параметров с «впечённым» check-first-адаптером. Модель для image-text-to-text на базе transformers; главное изменение — при выключенном thinking она чаще вызывает поиск перед ответом на вопросы вида «is X still …?», а не отвечает по памяти.

Ключевые особенности

  • ORPO LoRA на B5-27B: 259 пар из конфигов Nachschlagen round2 и round1.
  • Адаптер вшит в веса: все не-LoRA-тензоры совпадают с B5, целевые отличаются на alpha/r · B·A.
  • На held-out B6 делает правильный первый ход 41/41 против 35/41 у B5; из «нужен поиск» — 10/10 против 6/10.
  • На harder-наборе — 37/37 против 34/37; «нужен поиск» — 10/10 против 7/10.
  • Токенизатор, chat template, vision и MTP-веса взяты у B5 без изменений. Сделана для egirl; другие harness’ы не тестировались.

Пример

В карточке приводят пример устаревшего ответа B5: «Pat Gelsinger has been Intel's CEO since February 2021 and is still in the role as of today». B6 при выключенном thinking на вопросы «is X still …?» сначала вызывает поиск; если инструмент не нужен — отвечает сразу.

ПодробнееБенчмарки: ARC-Challenge 66,2 vs 65,9; ARC-Easy 87,2 vs 87,2; TruthfulQA 40,8 vs 40,9 — общие метрики остались на уровне.
Обучение: r16, alpha 32, lr 2e-5, beta 0.1, 2 эпохи, effective batch 8.
Ограничение: поиск при выключенном thinking добавляет несколько секунд; модель сделана под egirl, другие harness’ы не тестировались.
Запускали B5 и B6 как Q8_0 GGUF (B6 — B5 + f16-адаптер) в одной сессии.
Discuss Telegram 0 22

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 52 GB · ready files57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 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.

    # schneewolflabs/B6-27B
    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 schneewolflabs/B6-27B --local-dir ~/models/B6-27B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to B6-27B
  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

  1. B1-27Bfamily root
  2. B2-27B
  3. B3-27B
  4. B4-27B
  5. B5-27B
  6. B6-27B

New posts about versions of B1-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 1All on Hugging Face

About the model

schneewolflabs/B6-27B
Author
schneewolflabs
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

220

Ask AI about this model

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

Comments

Computer vision

CoSTA-MIX

CoSTA-MIX