ZatFung ยท Qwen3.8-Flash-Next๏ผˆไธญๆ–‡ + ไปฃ็ ๏ผ‰๏ผšSwift-1.5 ไบ”ๆกฃ + Unsloth Dynamic 3.1 320E

ๆœฌไป“ๅบ“ๆœ‰ไธคไธชๆฅๆบใ€ไธค็งๅ–่ˆ๏ผŒZatFung-โ€ฆ ๅ‰็ผ€็š„ๆฏไธชๅญ็›ฎๅฝ•้ƒฝๆ˜ฏไธ€ไธช็‹ฌ็ซ‹ๅฏ็”จ็š„ๆจกๅž‹๏ผš

็ณปๅˆ— ไธŠๆธธ ไธ“ๅฎถๆกฃ ๅญ็›ฎๅฝ•
Swift-1.5๏ผˆIQ3_S๏ผ‰ SC117/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF 192 / 256 / 288 / 320 / 384 5
Unsloth Dynamic 3.1๏ผˆUD-IQ3_XXS๏ผ‰ SC117/Qwen3.8-Flash-Next-abliterated-UD-GGUF 320 1๏ผˆๆŽจ่๏ผ‰

Swift ็ณปๅˆ—ๆ˜ฏ IQ3_S ๅŸบ็ก€้‡ๅŒ–๏ผŒๅœจๅŽŸๅง‹ 512 ไธชไธ“ๅฎถไธญไฟ็•™ 192 / 256 / 288 / 320 / 384 ไธช๏ผŒๅ…ฑไบ”ๆกฃ๏ผ› ไฟ็•™้ƒจๅˆ†็š„ๆƒ้‡ไฟๆŒๅŽŸๅง‹้‡ๅŒ–็ฒพๅบฆ๏ผŒๆœช้‡ๆ–ฐ้‡ๅŒ–ใ€‚UD ็ณปๅˆ—ๆ˜ฏ้€ๅฑ‚ๅŠจๆ€้‡ๅŒ–๏ผŒๅŒไฝ“็งฏไธ‹่ดจ้‡ๆ›ดๅฅฝใ€ ้€Ÿๅบฆ็จๆ…ขโ€”โ€”่งไธ‹้ข็š„ไธ“่Š‚ใ€‚

ๆœฌไป“ๅบ“ๆจกๅž‹็š„ไผ˜ๅŠฟๆ˜ฏใ€Œไธญๆ–‡ + ไปฃ็ ใ€๏ผšๆœฌ็‰ˆๆœฌ้’ˆๅฏนไธญๆ–‡้•ฟๆ–‡ใ€็œŸๅฎžไปฃ็ ๏ผˆZig / Go / Rust / C++๏ผ‰ไธŽ ็ฎ—ๆณ•้ข˜่ฟ™ไธ‰็ฑป่ดŸ่ฝฝๅšไบ†ๅ–่ˆ๏ผŒๅ› ๆญคๅœจ่ฟ™ไธค็ฑปๅœบๆ™ฏไธ‹๏ผŒๅŒไฝ“็งฏ็š„่กจ็Žฐๆ˜Žๆ˜พๅฅฝไบŽๆŒ‰้€š็”จ่ฏญๆ–™ๅ–่ˆ็š„็‰ˆๆœฌใ€‚ ้€‚ๅˆไปฃ็ ็ผ–็จ‹๏ผˆไป“ๅบ“็บงไปปๅŠกใ€ๅคš่ฏญ่จ€ใ€็ฎ—ๆณ•้ข˜๏ผ‰ไธŽๅฐ่ฏดๅˆ›ไฝœ๏ผˆ้•ฟไธญๆ–‡ใ€้œ€่ฆ็จณๅฎšๆ–‡้ฃŽไธŽ่ฟž่ดฏไธŠไธ‹ๆ–‡๏ผ‰ใ€‚

ไบ”ๆกฃ่ง„ๆ ผ

็›ฎๅฝ• ไธ“ๅฎถๆ•ฐ ็•™ๅ‡บ้›†ๅๅทฎ โ†“ ไธ‰็‰‡ๅˆ่ฎก ๅปบ่ฎฎ
ZatFung-Swift-384E-v5/ 384 0.88% 66.05 GiB ่ดจ้‡ไผ˜ๅ…ˆ
ZatFung-Swift-320E-v5/ 320 2.27% 60.16 GiB ๆŽจ่๏ผš่ดจ้‡/ไฝ“็งฏๅนณ่กก
ZatFung-Swift-288E-v5/ 288 3.18% 57.22 GiB ๅฐๆ˜พๅญ˜ / ๅฐๅ†…ๅญ˜
ZatFung-Swift-256E-v5/ 256 4.55% 54.27 GiB ๆž้™ๅŽ‹็ผฉ
ZatFung-Swift-192E-v5/ 192 8.59% 48.39 GiB ไป…ๅฎž้ชŒ๏ผŒไธๅปบ่ฎฎ็”Ÿไบง

ๅปบ่ฎฎไฝฟ็”จ 320 ๆกฃๅŠไปฅไธŠใ€‚ ใ€Œ็•™ๅ‡บ้›†ๅๅทฎใ€ๆ˜ฏๅœจๆœชๅ‚ไธŽๅ–่ˆ็š„็•™ๅ‡บ้›†ไธŠ๏ผŒไธŽๅŽŸๅง‹ 512 ไธ“ๅฎถๆจกๅž‹็›ธๆฏ” ่พ“ๅ‡บ่ทฏๅพ„ๅ‘็”Ÿๆ”นๅ˜็š„ๆฏ”ไพ‹โ€”โ€”่ถŠไฝŽ่ถŠๆŽฅ่ฟ‘ๅŽŸๆจกๅž‹ใ€‚192 ๆกฃๅทฒๆŽฅ่ฟ‘ 9%๏ผŒๆ—ฅๅธธไฝฟ็”จไผšๆ˜Žๆ˜พๆ„Ÿๅˆฐๆމ็‚นใ€‚

Unsloth Dynamic 3.1 ็‰ˆ ยท ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5/๏ผˆๆŽจ่๏ผ‰

**ๅŸบไบŽ SC117/Qwen3.8-Flash-Next-abliterated-UD-GGUF ็š„ UD-IQ3_XXS**๏ผšๅŒๆ ทๆŠŠ 512 ไธชไธ“ๅฎถ็ ๅˆฐ 320 ไธช๏ผŒๆ–ฉๆމ็š„ๆ˜ฏ่ทฏ็”ฑๅฎฝๅบฆไธŽไธ“ๅฎถๅ—๏ผŒ็•™ไธ‹ๆฅ็š„ๅ— ้€ๅญ—่Š‚ๆฌ่ฟใ€ไธ้‡ๆ–ฐ้‡ๅŒ–๏ผˆไธŽ Swift ็ณปๅˆ—ๅŒไธ€ๅฅ—ๅšๆณ•๏ผŒๅชๆ˜ฏๆบๅคดๆขๆˆไบ† Unsloth ็š„้€ๅฑ‚ๅŠจๆ€้‡ๅŒ–็‰ˆๆœฌ๏ผ‰ใ€‚

ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5/
  ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5-00001-of-00002.gguf   26.83 GiB   per-layer token embedding ่กจ๏ผˆPLE๏ผ‰
  ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5-00002-of-00002.gguf   33.40 GiB   ๅ…ถไฝ™ 1223 ไธชๅผ ้‡
                                                        ๅˆ่ฎก        60.23 GiB
  • ไธค็‰‡ๅฟ…้กปไธ€้ฝไธ‹่ฝฝ๏ผ›ๅŠ ่ฝฝๆ—ถๅช็ป™็ฌฌไธ€็‰‡็š„่ทฏๅพ„๏ผŒ็ฌฌไบŒ็‰‡ๆŒ‰ -0000N-of-0000M ่‡ชๅŠจๆ‰พ้ฝใ€‚
  • ็ฌฌไธ€็‰‡้‡Œๅชๆœ‰ PLE ่กจไธ€ๅผ ๅผ ้‡๏ผˆsplit.no = 0๏ผŒ160 ร— 320001536๏ผ‰๏ผŒๆ‰€ไปฅๅฎƒ 26.83 GiB ๅดๅชๆœ‰ไธ€ๅผ ่กจ๏ผŒๆ˜ฏๆญฃๅธธ็š„ใ€‚
  • ่ฟ™ไธ€ๆกฃ็š„ไธ“ๅฎถๆ˜ฏ้€ๅฑ‚ๅŠจๆ€้‡ๅŒ–๏ผšไธๅŒๅฑ‚็š„ gate/up ๅˆ†ๅˆซๆ˜ฏ IQ2_S / IQ3_XXS / IQ3_S / IQ4_XS๏ผŒdown ๆ˜ฏ IQ4_NL / Q2_0๏ผŒๅ› ๆญคๅŒไฝ“็งฏไธ‹้‡ๅŒ–่ฏฏๅทฎๆ›ดๅฐ๏ผ›ไปฃไปทๆ˜ฏๅ†…ๆ ธ็ป„ๅˆๆ›ด็ขŽ๏ผŒ้ข„ๅกซไธŽ่งฃ็ ้ƒฝไผšๆฏ” Swift ็ณปๅˆ—ๆ…ขไธ€ไบ›ใ€‚

ๅฆ‚ไฝ•้€‰๏ผšๆˆ‘ไปฌ่‡ชๅทฑๅœจไธญๆ–‡้•ฟๆ–‡๏ผˆๅฐ่ฏด๏ผ‰ไธŽ็œŸๅฎžไปฃ็ ๏ผˆZig / Go / Rust / C++๏ผ‰ไธŠๅฎžๆต‹๏ผŒUD3.1 320E ็š„่ดจ้‡ๆ˜Žๆ˜พๅฅฝไบŽ Swift 320E๏ผŒไฝ†้€Ÿๅบฆๆ›ดๆ…ขใ€‚ๆ…ขๅ‡ ้ธ ๅคšๅฎŒๅ…จๅ–ๅ†ณไบŽไฝ ็š„ๆŽจ็†็ซฏใ€ไธŠไธ‹ๆ–‡้•ฟๅบฆใ€KV ่ฎพ็ฝฎไธŽๆ˜ฏๅฆๆŠ•ๆœบ่งฃ็ โ€”โ€”ๅปบ่ฎฎ่‡ชๅทฑๅฎžๆต‹ไธ€ๆฌก๏ผŒ่ฆ่ดจ้‡้€‰ UD๏ผŒ่ฆ้€Ÿๅบฆ้€‰ Swiftใ€‚

ๆ–‡ไปถ็ป“ๆž„

Swift ไบ”ๆกฃๆฏๆกฃไธ€ไธชๅญ็›ฎๅฝ•๏ผŒๅ†…้ƒจๆ˜ฏ 3 ไธชๅˆ†็‰‡๏ผˆllama.cpp gguf-split ๅ‘ฝๅ๏ผ‰๏ผ›UD ้‚ฃไธ€ๆกฃๆ˜ฏ 2 ไธชๅˆ†็‰‡๏ผŒ่งไธŠไธ€่Š‚๏ผš

ZatFung-Swift-320E-v5/
  ZatFung-Swift-320E-v5-00001-of-00003.gguf    10.01 GiB   ๅ…ƒๆ•ฐๆฎ + embedding / attention / SSM / router
  ZatFung-Swift-320E-v5-00002-of-00003.gguf    23.33 GiB   ็ฌฌ 14โ€“47 ๅฑ‚็š„ไธ“ๅฎถไธŽๅ…ถไฝ™ๆƒ้‡
  ZatFung-Swift-320E-v5-00003-of-00003.gguf    26.82 GiB   per-layer token embedding ่กจ๏ผˆPLE / n-gram ่กจ๏ผ‰
  • ็ฌฌ 3 ็‰‡๏ผˆper_layer_token_embd.weight๏ผ‰ไบ”ๆกฃ้ƒฝไธ€ๆ ทๅคง๏ผˆ26.82 GiB๏ผ‰๏ผŒไธ้šไธ“ๅฎถๆ•ฐๅ˜ๅŒ–โ€”โ€” ๆ‰€ไปฅๆœ€ๅฐๆกฃไนŸๆœ‰ 48 GiBใ€‚
  • ๆฏ็‰‡ๅธฆ split.count / split.no / split.tensors.count๏ผŒๅŒไธ€ๆกฃ็š„ไธ‰็‰‡ๅฟ…้กปไธ€้ฝไธ‹่ฝฝใ€‚
  • ๅŠ ่ฝฝๆ—ถๅช็ป™็ฌฌไธ€็‰‡็š„่ทฏๅพ„๏ผŒๅ…ถไฝ™ๅˆ†็‰‡ๆŒ‰ๆ–‡ไปถๅ่‡ชๅŠจๆ‰พ้ฝ๏ผ›็ผบ็‰‡ไผšๆ˜Ž็กฎๆŠฅๅ‡บ็ผบๅ“ชไธ€็‰‡ใ€‚

้…ๅฅ—ๆ–‡ไปถ๏ผˆไป“ๅบ“ๆ น๏ผ‰

ๆ–‡ไปถ ๅคงๅฐ ็”จ้€”
mmproj-Qwen3.8-Flash-Next-BF16.gguf 0.85 GiB ่ง†่ง‰็ผ–็ ๅ™จ๏ผˆBF16๏ผ‰ใ€‚่ฆๅšๅ›พๅƒ่พ“ๅ…ฅๆ—ถไธŽไธปๆจกๅž‹้…ๅฏนไฝฟ็”จ
tmpls/qwen3.8-froggeric-v22.5.jinja 28 KB ่Šๅคฉๆจกๆฟ๏ผŒๅŽŸ็‰ˆ
tmpls/qwen3.8-froggeric-v22.5-hont.jinja 28 KB ๅŒไธŠ๏ผŒๆ€่€ƒๅ‰็ผ€็”จ็นไฝ“ใ€Œๆ€่€ƒ้Ž็จ‹๏ผšใ€
tmpls/qwen3.8-froggeric-v22.5-hons.jinja 28 KB ๅŒไธŠ๏ผŒๆ€่€ƒๅ‰็ผ€็”จ็ฎ€ไฝ“ใ€Œๆ€่€ƒ่ฟ‡็จ‹๏ผšใ€
mtp/*.gguf 0.4โ€“3.9 GiB ๅฏ้€‰็š„ MTP draft ๅคด๏ผˆ5 ไธช๏ผŒ้‡ๅŒ–ไธŽไธ“ๅฎถๆ•ฐไธๅŒ๏ผ‰ใ€‚ไธŽไธปๆจกๅž‹ๅฎŒๅ…จ็‹ฌ็ซ‹๏ผŒๅชๆœ‰ๅšๆŠ•ๆœบ่งฃ็ ๅŠ ้€Ÿๆ—ถๆ‰้œ€่ฆ
mtp/rt/ 787 MB ๆŸไธช่‰็จฟๅคด็š„่ฟ่กŒๆ—ถ็›ฎๅฝ•ๅฝขๆ€๏ผˆdense.bin + experts.bin + draft_vocab.bin + dense.txt๏ผ‰๏ผŒไพ› STRATA ๅผ•ๆ“Ž็›ดๆŽฅๅŠ ่ฝฝ๏ผŒไธๅฟ…ๅ†ไปŽ GGUF ่ฝฌๆข

ๆ‰€ๆœ‰ๆกฃไฝ๏ผˆSwift ไบ”ๆกฃไธŽ UD ไธ€ๆกฃ๏ผ‰้ƒฝไธๅŒ…ๅซ่ฟ™ไบ›ๆ–‡ไปถ๏ผŒๆŒ‰้œ€ๅ•็‹ฌไธ‹่ฝฝใ€‚

ๅฆ‚ไฝ•่ท‘

่ฟ™ๆ˜ฏๆ ‡ๅ‡† GGUF ๅˆ†็‰‡๏ผŒ้œ€่ฆๆ”ฏๆŒ qwen4exp๏ผˆQwen3.8-Flash-Next๏ผ‰ๆžถๆž„็š„ๆŽจ็†็ซฏๅŠ ่ฝฝ๏ผš

# Swift ๆกฃ๏ผˆ3 ็‰‡๏ผ‰
<ไฝ ็š„ๆŽจ็†็ซฏ> --model ZatFung-Swift-320E-v5/ZatFung-Swift-320E-v5-00001-of-00003.gguf --ctx 131072

# UD ๆกฃ๏ผˆ2 ็‰‡๏ผŒๆŽจ่๏ผ‰
<ไฝ ็š„ๆŽจ็†็ซฏ> --model ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5/ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5-00001-of-00002.gguf --ctx 131072

ไธŠไธ‹ๆ–‡ๅŽŸ็”Ÿๆ”ฏๆŒ 262144๏ผ›ๆ˜พๅญ˜/ๅ†…ๅญ˜ไธ่ถณๆ—ถไผ˜ๅ…ˆ้™ไธŠไธ‹ๆ–‡ๆˆ–้€‰ๆ›ดๅฐ็š„ๆกฃ๏ผŒ่€Œไธๆ˜ฏๆŠŠๆƒ้‡ๅ†้‡ๅŒ–ไธ€ๆฌกใ€‚

่ฎธๅฏ

่ฎธๅฏ็ปงๆ‰ฟไธŠๆธธ๏ผšSwift ๆกฃ็š„ๅŸบ็ก€ๆจกๅž‹ๆฅ่‡ช SC117 ็š„ Swift-1.5-Qwen3.8-Flash-Next๏ผˆGSQ-RCO abliterated๏ผ‰๏ผ› UD ๆกฃๆฅ่‡ช SC117 ็š„ Qwen3.8-Flash-Next-abliterated-UD-GGUF๏ผˆUnsloth Dynamic ้‡ๅŒ–๏ผ‰ใ€‚ไธค่€…็š„ๅบ•ๅฑ‚้ƒฝๆ˜ฏๅƒ้—ฎ ๏ผˆQwen๏ผ‰ๅŸบ็ก€ๆจกๅž‹๏ผŒๅ› ๆญคๆœฌไป“ๅบ“ๆฒฟ็”จๅฎƒไธŠๆธธ็š„่ฎธๅฏไธŽไฝฟ็”จๆกๆฌพ๏ผ›abliterated ไธŽๅŠจๆ€้‡ๅŒ–้ƒจๅˆ†็š„่ฏดๆ˜ŽไปฅไธŠๆธธไป“ๅบ“ไธบๅ‡†ใ€‚ ๅ•†็”จๅ‰่ฏท่‡ช่กŒๆ ธๅฏนไธŠๆธธ่ฎธๅฏๅŽŸๆ–‡ใ€‚

English (short)

Expert-reduced variants of SC117/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF (IQ3_S), keeping 192 / 256 / 288 / 320 / 384 of the original 512 experts. Weights keep their original quantisation (nothing is re-quantised). This release is tuned for Chinese long-form text, real code (Zig / Go / Rust / C++) and algorithm tasks โ€” hence the "Chinese + code" strength, aimed at coding and long-form Chinese fiction. Each tier is 3 gguf-split shards; the third shard is the 26.82 GiB per-layer token embedding table and is identical across tiers. Use 320E or 384E; 192E/256E deviate 8.6% / 4.6% from the original model and are not recommended for production. Licence follows the upstream Qwen-based model.

The second family, ZatFung-Unsloth-Dynamic-3-1-IQ3_XXS-320E-v5/, is the same 320-expert reduction of SC117/Qwen3.8-Flash-Next-abliterated-UD-GGUF (Unsloth Dynamic UD-IQ3_XXS, per-layer mixed quantisation) in 2 shards (shard 1 is the PLE table alone). We measured it to be better quality than the Swift 320E and somewhat slower on Chinese long-form and real code; pick it for quality, the Swift tiers for speed.

Downloads last month
1,761
GGUF
Model size
101B params
Architecture
qwen4exp
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

4-bit

8-bit

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

Model tree for NaamFung/ZatFung