Qwen3.8-Flash-Next-Uncensored ยท MTPLX Optimized Speed

English summary: This is an Apple Silicon (MLX/MTPLX) quantization of orcarouter/Qwen3.8-Flash-Next-Uncensored โ€” OrcaRouter's abliterated (refusal-removed) BF16 build of Qwen's Qwen3.8-Flash-Next (125B-A6B MoE with a native MTP speculative-decoding head). Quantization was performed locally with MTPLX Forge using the official flash-next-optimized-speed recipe (4-bit body + 8-bit sparse-attention projections), the same recipe as Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed. The native MTP head is preserved and verified: 2.07x faster than plain decoding in Forge verification on an M5 Max 128 GB. Requires MTPLX on Apple Silicon (96โ€“128 GB+ unified memory).


่ฟ™ๆ˜ฏไป€ไนˆ

่ฟ™ๆ˜ฏ OrcaRouter ๅŽปๅฎกๆŸฅๆƒ้‡็š„ MTPLX ้‡ๅŒ–็‰ˆใ€‚ๅ…ทไฝ“ๆฅ่ฏด๏ผš

  • ๆƒ้‡ๆฅๆบ๏ผšorcarouter/Qwen3.8-Flash-Next-Uncensored โ€”โ€” OrcaRouter ๅฏน Qwen/Qwen3.8-Flash-Next๏ผˆ125B ๅ‚ๆ•ฐ / ็บฆ 6B ๆฟ€ๆดป็š„ MoE๏ผŒๆททๅˆๆณจๆ„ๅŠ›๏ผŒๅŽŸ็”Ÿ่ง†่ง‰๏ผŒ่‡ชๅธฆ MTP ๆŠ•ๆœบ่งฃ็ ๅคด๏ผ‰ๅš็š„ๆถˆ่žๅผๅŽปๅฎกๆŸฅ๏ผˆabliteration๏ผ‰BF16 ็‰ˆๆœฌ
  • ้‡ๅŒ–ๅˆถไฝœ๏ผšๆœฌไป“ๅบ“็š„้‡ๅŒ–ๅนถ้ž Orca ๅฎ˜ๆ–นๅ‘ๅธƒ๏ผŒ่€Œๆ˜ฏไฝฟ็”จ MTPLX Forge๏ผˆv2.12.2๏ผ‰ๅœจๆœฌๅœฐไปŽไธŠ่ฟฐ BF16 ๆƒ้‡ๆž„ๅปบ๏ผŒ้‡‡็”จ MTPLX ๅฎ˜ๆ–นๅ†…็ฝฎ้…ๆ–น flash-next-optimized-speed โ€”โ€” ไธŽ Youssofal ๅฎ˜ๆ–นๅŒ…ๅฎŒๅ…จ็›ธๅŒ็š„้…ๆ–น
  • ่ฟ่กŒ็Žฏๅขƒ๏ผšMTPLX๏ผˆApple Silicon ไธ“ๅฑž๏ผŒmacOS 14+๏ผ‰๏ผŒ ๅˆฉ็”จๆจกๅž‹่‡ชๅธฆ็š„ MTP ๅคดๅš็ฒพ็กฎๆŠ•ๆœบ้‡‡ๆ ทๅŠ ้€Ÿ

้‡ๅŒ–้…็ฝฎ๏ผˆq4 + q8 ๆณจๆ„ๅŠ›๏ผ‰

็ป„ไปถ ็ฒพๅบฆ ่ฏดๆ˜Ž
ๆจกๅž‹ไธปไฝ“๏ผˆไธ“ๅฎถใ€ๆŠ•ๅฝฑ็ญ‰๏ผ‰ 4-bit affine, group 32 ๅŠจๆ€ๆททๅˆ็ฒพๅบฆ
็จ€็–ๆณจๆ„ๅŠ›ๆŠ•ๅฝฑ๏ผˆQSA๏ผ‰ 8-bit, group 64 ไฟ่ดจ้‡็š„ๅ…ณ้”ฎ้ƒจไฝ
MTP ๆŠ•ๆœบ่งฃ็ ๅคด 4-bit, group 64 ๅŽŸ็”ŸๅคดๅฎŒๆ•ดไฟ็•™
n-gram ๆŸฅ่ฏ่กจ 4-bit, group 32๏ผˆ29.8 GiB๏ผ‰ ่ฟ่กŒๆ—ถไปŽ SSD ๆตๅผ่ฏปๅ–๏ผŒไธๅ ๅ†…ๅญ˜
่ง†่ง‰ๅก” ๅฎŒๆ•ดไฟ็•™ 333 ไธชๅผ ้‡
lm_head / embed_tokens 8-bit, group 64
  • ๆ€ปๅคงๅฐ๏ผš115.3 GB๏ผ›ๆœ‰ๆ•ˆ็ฒพๅบฆ็บฆ 5.14 bit/ๆƒ้‡
  • ๅธธ้ฉปๅ†…ๅญ˜็บฆ 83 GiB๏ผŒๅฎ˜ๆ–นๅปบ่ฎฎ 128 GB ็ปŸไธ€ๅ†…ๅญ˜็š„ Mac๏ผˆๆœ€ไฝŽ 96 GB ๆกฃไฝ่ฏท็”จ Bare Speed ็ฑป้…ๆ–น๏ผ‰
  • ไธŠไธ‹ๆ–‡็ช—ๅฃ๏ผš262,144 tokens

ๅฎžๆต‹ๆ•ฐๆฎ๏ผˆM5 Max 128 GB๏ผŒMTPLX 2.12.2๏ผŒๆจกๅž‹ๅŽŸ็”Ÿ้‡‡ๆ ท่ฎพ็ฝฎ temp 1.0 / top-p 0.95 / top-k 20๏ผ‰

Forge ๅฎ˜ๆ–น้ชŒ่ฏ๏ผˆๆž„ๅปบๆ—ถ่‡ชๅŠจๅฎŒๆˆ๏ผ‰

ๆจกๅผ ้€Ÿๅบฆ ๅ€็އ
ๆ™ฎ้€š่‡ชๅ›žๅฝ’่งฃ็ ๏ผˆdepth 0๏ผ‰ 41.3 tok/s 1.0x
MTP ๆŠ•ๆœบ่งฃ็ ๏ผˆdepth 3๏ผ‰ 85.7 tok/s 2.07x

้ชŒ่ฏ็ป“่ฎบ mtp_depth_wins๏ผ›MTP ๅฅ‘็บฆ็ป 32 ็ป„ๅ€™้€‰ๅฎžๆต‹ๆ กๅ‡†๏ผŒไฟ็•™ๅฎถๆ—้ป˜่ฎคๅฅ‘็บฆใ€‚

ๅฎž้™…ไฝฟ็”จๅœบๆ™ฏๅคๆต‹๏ผˆOpenAI ๅ…ผๅฎน API๏ผŒๅ…ณ้—ญๆ€่€ƒๆจกๅผ๏ผ‰

ๅœบๆ™ฏ ๆœฌๅŒ… ๅฎ˜ๆ–นๅŒ…๏ผˆๅฏน็…ง๏ผ‰ ่ฏดๆ˜Ž
็ผ–็จ‹ไปปๅŠก๏ผˆPython ็ผ“ๅญ˜็ฑป + ๅ•ๅ…ƒๆต‹่ฏ•๏ผ‰ 52โ€“71 tok/s 85 tok/s ่‰็จฟๅ‘ฝไธญ็އ็บฆ 92%/86%/81%๏ผŒไธŽๅฎ˜ๆ–นๅŒ…ๅ‡ ไนŽไธ€่‡ด๏ผ›ๆ ก้ชŒ่ทฏๅพ„็•ฅๆ…ขไบŽๅฎ˜ๆ–น็›–็ซ ็‰ˆ
ไธญๆ–‡้•ฟๆ–‡ๅ†™ไฝœ๏ผˆ2500 ๅญ—ๆ–‡็ซ ๏ผ‰ 26โ€“34 tok/s 32 tok/s ๅˆ›ๆ„ๅ†™ไฝœไธๅฏ้ข„ๆต‹ๆ€ง้ซ˜๏ผŒๅ‘ฝไธญ็އๅคฉ็„ถๅไฝŽ๏ผˆไธคๅŒ…่กจ็Žฐไธ€่‡ด๏ผ‰๏ผ›้™ไฝŽๆธฉๅบฆๅฏๆ˜Žๆ˜พๆ้€Ÿ
้ฆ– token ๅปถ่ฟŸ 0.22โ€“0.29 s 0.22 s ๅฐๆ็คบ่ฏไธ‹ๅŸบๆœฌ็žฌๆ—ถ

ๆœฌๆœบๆทฑๅบฆ่ฐƒไผ˜๏ผˆmtplx tune --retune๏ผŒM5 Max 128 GB๏ผ‰

ๆจกๅผ ้€Ÿๅบฆ ็›ธๅฏน่ฃธ่งฃ็ 
AR๏ผˆๆ— ๆŠ•ๆœบ่งฃ็ ๏ผ‰ 42.0 tok/s 1.00x
D1 56.0 tok/s 1.33x
D2 40.8 tok/s 0.97x๏ผˆๆฏ”่ฃธ่งฃ็ ่ฟ˜ๆ…ข๏ผ‰
D3 66.4 tok/s 1.58x๏ผŒๆœฌๆœบๆœ€ไผ˜

็ป“่ฎบ๏ผš็ผ–็จ‹ / ็ป“ๆž„ๅŒ–่พ“ๅ‡บๅœบๆ™ฏๆ้€Ÿๆ˜พ่‘—๏ผˆ็›ธๅฏน่ฃธ่งฃ็  1.3โ€“2.1 ๅ€๏ผ‰๏ผ›้ซ˜ๆธฉๅˆ›ๆ„ๅ†™ไฝœๅœบๆ™ฏๅปบ่ฎฎ่ฐƒไฝŽ temperature ๏ผˆๅฆ‚ 0.6โ€“0.7๏ผ‰ไปฅๆ้ซ˜่‰็จฟๅ‘ฝไธญ็އใ€‚

ไฝฟ็”จๆ–นๆณ•

# ๅฎ‰่ฃ… MTPLX๏ผˆ้œ€่ฆ Apple Silicon + macOS 14+๏ผ‰
brew install youssofal/mtplx/mtplx

# ๆ‹‰ๅ–ๅนถ่ฟ่กŒๆœฌๆจกๅž‹
mtplx pull NeosQ/Qwen3.8-Flash-Next-Uncensored-MTPLX-Optimized-Speed
mtplx serve --model NeosQ/Qwen3.8-Flash-Next-Uncensored-MTPLX-Optimized-Speed

ๆœๅŠกๅฏๅŠจๅŽๅœจ http://127.0.0.1:8000 ๆไพ› OpenAI / Anthropic ๅ…ผๅฎน API๏ผŒๅฏ็›ดๆŽฅๆŽฅ Claude Codeใ€Clineใ€ OpenCode ็ญ‰ๅฎขๆˆท็ซฏใ€‚ไนŸๅฏไปฅ็›ดๆŽฅไฝฟ็”จ MTPLX ๆกŒ้ขๅบ”็”จ๏ผŒๅœจๆจกๅž‹ๅˆ—่กจไธญ้€‰ๆ‹ฉๅณๅฏใ€‚

โš ๏ธ ๅ…่ดฃๅฃฐๆ˜Ž๏ผˆๆฒฟ็”จ่‡ชไธŠๆธธ๏ผ‰

ๆœฌๆจกๅž‹็š„ไธŠๆธธๆƒ้‡็ป่ฟ‡ไบ†ๅŽปๅฎกๆŸฅๅค„็†๏ผˆabliteration๏ผ‰๏ผŒๅ…ถๅฎ‰ๅ…จๅฏน้ฝๅทฒ่ขซๅคงๅน…็งป้™ค๏ผš

  • ๅฎƒไผšๅ“ๅบ”ๅŽŸ็‰ˆ Qwen3.8-Flash-Next ๆ‹’็ป็š„ๆœ‰ๅฎณใ€ไธ้“ๅพทๆˆ–้žๆณ•่ฏทๆฑ‚๏ผŒๆฒกๆœ‰ๅ†…็ฝฎๆŠคๆ 
  • ไป…ไพ›ๅˆๆณ•็ ”็ฉถ็”จ้€”๏ผˆๅฏ่งฃ้‡Šๆ€งใ€AI ๅฎ‰ๅ…จไธŽๆ‹’็ญ”ๆœบๅˆถ็ ”็ฉถใ€็บข้˜Ÿๆต‹่ฏ•ใ€ๅ—ๆŽงๅฎž้ชŒ๏ผ‰
  • ไฝฟ็”จ่€…ๅฏนไฝฟ็”จๆ–นๅผๅŠ็”Ÿๆˆๅ†…ๅฎนๆ‰ฟๆ‹…ๅ…จ้ƒจ่ดฃไปป๏ผ›่ฏทๅ‹ฟๅœจๆœชๆทปๅŠ ่‡ชๆœ‰ๅฎ‰ๅ…จๅฎกๆ ธๅฑ‚็š„ๆƒ…ๅ†ตไธ‹้ƒจ็ฝฒ็ป™็ปˆ็ซฏ็”จๆˆท
  • ไฝฟ็”จ้กป้ตๅฎˆ็ปงๆ‰ฟ่‡ชๅŸบๅบงๆจกๅž‹็š„ Apache 2.0 ่ฎธๅฏ่ฏๅŠๆ‰€ๆœ‰้€‚็”จๆณ•ๅพ‹ๆณ•่ง„

่ฎธๅฏไธŽ่‡ด่ฐข

  • License: Apache 2.0๏ผˆ็ปงๆ‰ฟ่‡ช Qwen/Qwen3.8-Flash-Next๏ผ‰
  • ๅŽปๅฎกๆŸฅๆƒ้‡๏ผšOrcaRouter๏ผˆorcarouter/Qwen3.8-Flash-Next-Uncensored๏ผ‰
  • ่ฟ่กŒๆ—ถไธŽ้‡ๅŒ–้…ๆ–น๏ผšMTPLX by Youssof Altoukhi๏ผˆgithub.com/youssofal/MTPLX๏ผŒApache-2.0๏ผ‰
  • ๅŸบๅบงๆจกๅž‹๏ผšQwen / Alibaba
Downloads last month
1,032
Safetensors
Model size
25B params
Tensor type
U32
ยท
BF16
ยท
I64
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

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