Skip to content
Language models

K2H S1 final

MoE-модель для кода и ML-инфраструктуры

1 min readupdated

K2H S1 final — это чекпоинт после второго этапа дообучения IFM/K2-Horizon-MoVA-36B-A4B: 36B параметров всего и около 4B активных, архитектура MoE. Модель специализировали под LLM-инференс и инженерию ML-систем, сохранив общие навыки, следование инструкциям и tool calling; основной прирост — в коде.

Ключевые особенности

  • База — K2-Horizon-MoVA-36B-A4B; те же архитектура, токенизатор и chat template.
  • Заметно лучше решает задачи на код: HumanEval 68.9 против 53.1, MBPP 72.2 против 65.2 у релиза.
  • Tool calling удержан: добавлен якорь на 8 787 ответах, чтобы компенсировать просадку BFCL.
  • Длинный контекст до 524 288 токенов; на 32k–64k возможна небольшая просадка, на 128k — выше по точечной оценке.
  • Веса в bf16 safetensors, загрузка с trust_remote_code=True.

Пример

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Avifenesh/k2h-s1-final"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype="bfloat16")
inputs = tok("Explain MoE inference.", return_tensors="pt")
out = model.generate(**inputs, temperature=1.0, top_p=0.95)
ПодробнееБенчмарки: IFEval 84.9 против 83.7, MMLU 77.3 без изменений, BFCL немного ниже, но разница незначима.
Длинный контекст: RULER 32k — 86.8, 64k — 81.3, 128k — 81.6; на 32k–64k возможна просадка в несколько пунктов.
Few-shot GSM8K просел на 2.3 пункта, chat GSM8K и MATH-500 держатся.
Обучение: 150 шагов, около 118M токенов, 3 GPU, bf16 и FSDP2.
Discuss Telegram 0 418

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 70 GB · ready files72 GB
  • FP8 / INT8recommendedweights 35 GB37 GB
  • 4-bitweights 20 GB22 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.

Language models: with a 8192-token context.

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

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent RTX A6000

    $0.43 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # Avifenesh/k2h-s1-final — RTX A6000, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve Avifenesh/k2h-s1-final --max-model-len 8192 --quantization fp8
    Add more LoRAs to k2h-s1-final
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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 K2-Horizon-MoVA-36B-A4B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

Avifenesh/k2h-s1-final
Author
Avifenesh
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

4180

Ask AI about this model

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

Comments