Skip to content
Language models

Kimi-K3.1-Dflash2-Preview

драфт для ускорения Kimi-K3.1-Pro

2 min read

Релиз RadixArk: DFlash2-спекулятор для Kimi-K3.1-Pro, который ускоряет инференс через speculative decoding в SGLang. Это не самостоятельная LLM, а BF16-драфт на 2,78 млрд параметров: пятислойный MLA-стек работает вместе с target-моделью и использует её эмбеддинги, LM head, tokenizer и chat template.

Ключевые особенности

  • Пятислойный MLA-драфт: hidden size 7168, 96 heads, q LoRA rank 1536, KV LoRA rank 512.
  • Читает hidden states из шести слоёв target [7, 23, 51, 67, 83, 91], объединяет их одной проекцией и применяет block-causal sliding-window attention с окном 1 024 токена.
  • Low-rank селектор (rank 256, top-k 32) выбирает каждый предложенный токен из top-k кандидатов с учётом предыдущего токена — блок сэмплируется как путь, а не по позициям.
  • Блочные свёртки (kernel 2, group 16) в ветках attention и MLP каждого слоя.
  • Конфиг на ширину верификации 5 токенов (1 текущий + 4 драфта); те же веса работают и при ширине 8.
  • Контекст в конфиге — 1 048 576 токенов (YaRN, factor 32), как у target; это не то же самое, что окно attention.

Пример

python -m sglang.launch_server --model-path moonshotai/Kimi-K3.1-Pro --tp-size 8 --speculative-algorithm DFLASH --speculative-draft-model-path RadixArk/Kimi-K3.1-Dflash2-Preview --speculative-num-draft-tokens 5 --speculative-use-block-verification
Подробнееacc_len при ширине 5: GSM8K 4,259; AIME26 3,232; GPQA Diamond 3,276. При ширине 8: 5,780 / 3,696 / 3,768.
Замеры — на одном узле 8× B300 (TP8), FP8 KV cache для target и драфта, контекст 131 072, до 64 параллельных запросов при ширине 5.
Драфт не самостоятельная LLM; конфиг контекста 1M не гарантирует качество принятия на всей длине. В карточке измерена только acceptance length, не точность ответов.
Discuss Telegram 1

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # RadixArk/Kimi-K3.1-Dflash2-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 RadixArk/Kimi-K3.1-Dflash2-Preview --local-dir ~/models/Kimi-K3.1-Dflash2-Preview
    # The model is in ~/models. How to run it is in the model card (link above).
  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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of Kimi-K3.1-Dflash2-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

RadixArk/Kimi-K3.1-Dflash2-Preview
Author
RadixArk
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments