Qwen3.8-Flash-Next-MTP-bf16

Qwen/Qwen3.8-Flash-Next ใฎๅ…ฌๅผ bf16 ใƒใ‚งใƒƒใ‚ฏใƒใ‚คใƒณใƒˆใซๅ†…่”ตใ•ใ‚ŒใŸ MTP ใƒ˜ใƒƒใƒ‰๏ผˆmtp.* 31 ใƒ†ใƒณใ‚ฝใƒซ๏ผ‰ใ‚’ๅˆ‡ใ‚Šๅ‡บใ—ใŸใ€ ๆŠ•ๆฉŸใƒ‡ใ‚ณใƒผใƒ‰็”จใƒ‰ใƒฉใƒ•ใ‚ฟใƒผใ€‚ๅ˜ไฝ“ใงใฏไฝฟใ‚ใชใ„ใ€‚

  • ๅค‰ๆ›ๅ…ƒ: Qwen/Qwen3.8-Flash-Next๏ผˆrevision de4b8e4๏ผ‰
  • ๅค‰ๆ›ใƒ„ใƒผใƒซ: mlx-vlm main๏ผˆ0.7.0rc0๏ผ‰ใฎ Qwen4ExpMTPSplitter
  • ใ‚ตใ‚คใ‚บ: 4.9 GiB๏ผˆbf16๏ผ‰

ๅฟ…่ฆ็’ฐๅขƒ

qwen4_exp_mtp ใฎๆŠฝๅ‡บใƒปๅฎŸ่กŒใฏ mlx-vlm 0.7.0 ไปฅ้™๏ผˆใƒชใƒชใƒผใ‚น็‰ˆ 0.6.17 ใซใฏๆœชๅŽ้Œฒ๏ผ‰ใ€‚

pip install "mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm@main"

ไฝฟใ„ๆ–น

python -m mlx_vlm generate --model ToPo-ToPo/Qwen3.8-Flash-Next-mlx-4bit \
    --draft-model ToPo-ToPo/Qwen3.8-Flash-Next-MTP-bf16 \
    --draft-kind mtp --draft-block-size 2 \
    --prompt "..." --max-tokens 512

้‡ๅญๅŒ–็‰ˆใƒชใƒใ‚ธใƒˆใƒช๏ผˆ4bit / 8bit๏ผ‰ใฎ mtp/ ใ‚ตใƒ–ใƒ•ใ‚ฉใƒซใƒ€ใซใ‚‚ๅŒใ˜ใ‚‚ใฎใŒๅ…ฅใฃใฆใ„ใ‚‹ใฎใงใ€ๆœฌไฝ“ใ‚’่ฝใจใ›ใฐใƒ‰ใƒฉใƒ•ใ‚ฟใƒผใ‚‚ไป˜ใ„ใฆใใ‚‹ใ€‚

ๆŠฝๅ‡บใ‚ณใƒžใƒณใƒ‰

from mlx_vlm.speculative.drafters.qwen4_exp_mtp.split import Qwen4ExpMTPSplitter
Qwen4ExpMTPSplitter().split("Qwen/Qwen3.8-Flash-Next", "Qwen3.8-Flash-Next-MTP-bf16")

้‡ๅญๅŒ–ๅพŒใฎใƒชใƒใ‚ธใƒˆใƒชใ‹ใ‚‰ใฏ mtp.* ใŒ่ฝใกใฆใ„ใ‚‹๏ผˆmlx-vlm ใฎ sanitize ใŒ้™คๅค–ใ™ใ‚‹๏ผ‰ใฎใงใ€ ๅฟ…ใšๅ…ฌๅผ bf16 ใ‹ใ‚‰ๅˆ‡ใ‚Šๅ‡บใ™ใ“ใจใ€‚

Downloads last month
175
Safetensors
Model size
3B params
Tensor type
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ToPo-ToPo/Qwen3.8-Flash-Next-MTP-bf16

Finetuned
(141)
this model