Qwen3.8-Flash-Next IQ3_M GGUF (for a 24 GB GPU + 64 GB RAM PC) โ€” BETA

  • 24G64G = for a PC with a 24 GB GPU (VRAM) + 64 GB RAM - a target setup, not a guarantee of speed.
  • a mixed-type file - the experts are IQ3_S 57%, IQ4_NL 34%, Q8_0 7%, IQ4_XS 1% (by bytes), 3.94 bits a weight on average, about IQ3_M.

A GGUF of Qwen/Qwen3.8-Flash-Next made for one 24 GB GPU and 64 GB of RAM. Every expert is re-quantized with LDLQ from NVIDIA-format NVFP4 weights, in exactly unsloth UD-IQ4_XS's types and size (93.7 GB). All 512 experts are kept. Same size, same speed on the same emulated PC (69.9 vs 71.1 tok/s on coding tasks), and KL 0.068 vs 0.092 (-26%) against the NVFP4 model; KL on Japanese text 0.146 vs 0.195. Runs in Abyss and llama.cpp.

BETA: measured on one PC only (below). Reports from real 24 GB GPUs and other RAM sizes are very welcome (Community tab, or Abyss issues).

For a 16 GB GPU: Qwen3.8-Flash-Next-IQ3_M-16G64G-GGUF. For 12 GB + 32 GB: Qwen3.8-Flash-Next-E448-IQ2_M-12G32G-GGUF.

ๆ—ฅๆœฌ่ชžใฎ่ชฌๆ˜Žใฏไธ‹ใซใ‚ใ‚Šใพใ™ใ€‚

Pick a size (this model's other repos)

repo for a PC with download note
Qwen3.8-Flash-Next-E448-IQ2_M-12G32G-GGUF 12 GB GPU + 32 GB RAM 67.5 GB 448 of 512 experts kept
Qwen3.8-Flash-Next-IQ3_M-16G64G-GGUF 16 GB GPU + 64 GB RAM 86.9 GB all 512 experts
this repo 24 GB GPU + 64 GB RAM 93.7 GB all 512 experts, higher quality
Qwen3.8-Flash-Next-DGX-Spark-GGUF DGX Spark (128 GB) 86.9 GB same files as 16G64G

Earlier repo name: Qwen3.8-Flash-Next-24G64G-GGUF (old links redirect here).

Names: the quant in the repo name is the nearest llama.cpp type to the experts' average bits; most files mix several types (listed above), so they are not a stock quant of that name. E<n> = experts kept per layer (only when some were removed).

Compared with the same-size file

download resident (GPU + RAM) KL (10 seqs) same top token KL, Japanese KL, code decode, 3 coding tasks
this file 93.7 GB 64.2 GB 0.068 90.3% 0.146 0.092 69.9 tok/s
unsloth UD-IQ4_XS 93.7 GB 64.9 GB 0.092 88.7% 0.195 0.116 71.1 tok/s

KL against the NVFP4 model (10 held-out sequences x 1,023 tokens: code, English, Chinese, Japanese and others), Abyss engine. "Resident" leaves out the 28.8 GB per-layer embedding table, which is read a few rows a token from the SSD. The two files have the same tensor types; only the rounding (LDLQ, calibrated with Japanese and chat turns) differs.

Real use

Abyss server, thinking on, two seeds, on the emulated PC (below):

this file unsloth UD-IQ4_XS
Japanese (5 prompts: business email, explainer, summary, literature fact, arithmetic) 10/10 answered, facts and sums right 10/10
5 coding tasks with hidden tests 5/5, 4/5 3/5, 5/5
8K-token prompt, find the needle OK / OK OK / OK
empty answers / prompt echoes none none
prompt processing (8K) 1,680-1,689 tok/s 1,670-1,683 tok/s

Speed

Abyss v1.3.108, a 24 GB GPU + 64 GB RAM PC emulated on an RTX 5090 (32 GB): VRAM capped at 24 GB (STRATA_GLM_VRAM_GB=24), RAM tier 46 GB inside a 58 GB cgroup, 8 CPU threads, iGPU off, NVMe SSD, MTP drafter on.

  • 3 coding tasks: 69.9 tok/s (unsloth UD-IQ4_XS 71.1)
  • Japanese, 3 prompts: 61-77 tok/s
  • llama.cpp, this file, --n-cpu-moe 34 (24.2 GB of VRAM): 26-34 tok/s

Abyss v1.3.108 or later is needed for the prompt path: the Q8_0 down experts of this layout read past their slot in older versions (also in unsloth's UD-IQ4_XS).

Other GPUs are not measured with this file. Reports welcome.

Files

Qwen3.8-Flash-Next-24G64G-0000N-of-00003.gguf, 3 shards; point your engine at -00001-of-00003.gguf. SHA256SUMS in the repo. Experts: gate/up IQ3_S (layer 2 IQ4_XS), down IQ4_NL (layers 2, 4, 30, 46, 47 Q8_0). Non-expert tensors (attention, router, shared expert, embeddings, per-layer embedding table) are unsloth UD-IQ4_XS's bytes, unchanged.

Feedback (this is a BETA)

  • Speed on your PC: ./abyss.sh --bench --deep in Abyss, post abyss-bench.txt with this issue form (change the title to your GPU), with your RAM size and SSD.
  • Quality in real use: empty answers, wrong language, loops, places where it is weaker than unsloth's file. A prompt that reproduces it helps most.

How it was made

With trench-gguf (tools/hessian_dump.cpp, tools/requant_experts.py, tools/merge_experts.py):

  1. Hessians: the NVFP4 model (magnitudedev/Qwen3.8-Flash-Next-NVFP4-GGUF) over 1.26M tokens: its own chat turns (181 prompts: Japanese 44, Chinese 20, English 20, code 37, math 35, tool calls 25) followed by mixed text in the same languages and domains.
  2. LDLQ: every expert re-quantized from the NVFP4 weights (per-expert scales folded in) in UD-IQ4_XS's types, then merged into unsloth UD-IQ4_XS.

Running it

  • Abyss v1.3.108 or later. Pack it once:
    cd Abyss
    .venv/bin/python tools/iq_pack.py --fast --gguf /path/to/Qwen3.8-Flash-Next-24G64G-00001-of-00003.gguf --out /path/to/pack-fast
    
    and start it with --glm-pack /path/to/pack-fast (STRATA_GLM_RAM_GB=46 on a 64 GB PC; STRATA_GLM_VRAM_GB=24 on a larger card to test the 24 GB case).
  • llama.cpp (arch qwen4exp, upstream ac8e1694d of 2026-10-29): runs as is. Checked with llama-server -m Qwen3.8-Flash-Next-24G64G-00001-of-00003.gguf -ngl 99 --n-cpu-moe 34 -c 32768 --jinja -t 8 (24.2 GB of VRAM; use 35 or more on a real 24 GB card), the same real-use prompts: Japanese 5/5 answered (none empty), coding 3/5, the 8K needle found; decode 26-34 tok/s.

License

The original model's license (qwen-community-1.0). All credit for the model goes to the Qwen team; the NVFP4 weights the experts were made from are magnitudedev's, the non-expert tensors unsloth's.


ๆ—ฅๆœฌ่ชž

  • 24G64G = GPU๏ผˆVRAM๏ผ‰24 GB + RAM 64 GB ใฎ PC ๅ‘ใ‘๏ผˆๆƒณๅฎšใ—ใฆใ„ใ‚‹ๆง‹ๆˆใงใ€้€Ÿใ•ใ‚’ไฟ่จผใ™ใ‚‹ใ‚‚ใฎใงใฏใ‚ใ‚Šใพใ›ใ‚“๏ผ‰ใ€‚
  • ๅฝขๅผใ‚’ๆททใœใŸใƒ•ใ‚กใ‚คใƒซใงใ™ใ€‚ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฏ IQ3_S 57%, IQ4_NL 34%, Q8_0 7%, IQ4_XS 1%๏ผˆๅฎน้‡ใฎๅ‰ฒๅˆ๏ผ‰ใ€ๅนณๅ‡ 3.94 ใƒ“ใƒƒใƒˆใงใ€IQ3_M ใใ‚‰ใ„ใงใ™ใ€‚

ใ‚ตใ‚คใ‚บใฎ้ธใณๆ–น๏ผˆๅŒใ˜ใƒขใƒ‡ใƒซใฎใปใ‹ใฎ็‰ˆ๏ผ‰

ใƒชใƒใ‚ธใƒˆใƒช ๅ‘ใ‘ใฎ PC ใƒ€ใ‚ฆใƒณใƒญใƒผใƒ‰ ใƒกใƒข
Qwen3.8-Flash-Next-E448-IQ2_M-12G32G-GGUF GPU 12 GB + RAM 32 GB 67.5 GB ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆ 512 ๅ€‹ไธญ 448 ๅ€‹
Qwen3.8-Flash-Next-IQ3_M-16G64G-GGUF GPU 16 GB + RAM 64 GB 86.9 GB 512 ๅ€‹ใ™ในใฆ
ใ“ใฎ็‰ˆ GPU 24 GB + RAM 64 GB 93.7 GB 512 ๅ€‹ใ™ในใฆใ€ๅ“่ณช้‡่ฆ–
Qwen3.8-Flash-Next-DGX-Spark-GGUF DGX Spark๏ผˆ128 GB๏ผ‰ 86.9 GB 16G64G ใจๅŒใ˜ใƒ•ใ‚กใ‚คใƒซ

ไปฅๅ‰ใฎใƒชใƒใ‚ธใƒˆใƒชๅใฏ Qwen3.8-Flash-Next-24G64G-GGUF ใงใ™๏ผˆๅคใ„ URL ใฏใ“ใ“ใธ่ปข้€ใ•ใ‚Œใพใ™๏ผ‰ใ€‚

ๅๅ‰ใฎ้‡ๅญๅŒ–ใฎๅž‹ใฏใ€ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฎๅนณๅ‡ใƒ“ใƒƒใƒˆๆ•ฐใซใ„ใกใฐใ‚“่ฟ‘ใ„ llama.cpp ใฎๅž‹ใงใ™ใ€‚ๅคšใใฎใƒ•ใ‚กใ‚คใƒซใฏ่ค‡ๆ•ฐใฎๅž‹ใ‚’ๆททใœใฆใ„ใ‚‹ใฎใง๏ผˆไธŠใซๅ†…่จณ๏ผ‰ใ€ใใฎๅๅ‰ใฎๆ™ฎ้€šใฎ้‡ๅญๅŒ–ใจๅŒใ˜ใ‚‚ใฎใงใฏใ‚ใ‚Šใพใ›ใ‚“ใ€‚E<ๆ•ฐ> ใฏๅฑคใ”ใจใซๆฎ‹ใ—ใŸใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฎๆ•ฐใงใ™๏ผˆๆธ›ใ‚‰ใ—ใŸใจใใ ใ‘ไป˜ใ‘ใพใ™๏ผ‰ใ€‚

Qwen/Qwen3.8-Flash-Next ใฎ GGUF ใ‚’ใ€GPU 24 GB + RAM 64 GB ใฎ PC ๅ‘ใ‘ใซไฝœใ‚Šใพใ—ใŸใ€‚ ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฏใ€NVIDIA ๅฝขๅผใฎ NVFP4 ใฎ้‡ใฟใ‹ใ‚‰ LDLQ ใง้‡ๅญๅŒ–ใ—็›ดใ—ใฆใ„ใพใ™ใ€‚ๅž‹ใ‚‚ๅคงใใ•๏ผˆ93.7 GB๏ผ‰ใ‚‚ unsloth ใฎ UD-IQ4_XS ใจใพใฃใŸใๅŒใ˜ใงใ€ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฏ 512 ๅ€‹ใ™ในใฆๆฎ‹ใ—ใฆใ„ใพใ™ใ€‚ ๅŒใ˜่จญๅฎšใงๆฏ”ในใ‚‹ใจใ€้€Ÿใ•ใฏๅŒใ˜๏ผˆใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐใฎ็”Ÿๆˆ 69.9 ใจ 71.1 tok/s๏ผ‰ใงใ€ๅ…ƒใฎ NVFP4 ใฎใƒขใƒ‡ใƒซใจใฎๅทฎ๏ผˆKL๏ผ‰ใฏ 0.092 โ†’ 0.068๏ผˆ-26%๏ผ‰ใ€ๆ—ฅๆœฌ่ชžใฎๆ–‡็ซ ใงใฏ 0.195 โ†’ 0.146 ใงใ™ใ€‚Abyss ใจ llama.cpp ใงๅ‹•ใใพใ™ใ€‚

BETA ็‰ˆใงใ™ใ€‚ ๆธฌใฃใŸใฎใฏไธ‹ใฎ 1 ๅฐใ ใ‘ใงใ™ใ€‚ๆœฌ็‰ฉใฎ 24 GB ใฎ GPU ใ‚„ใ€ใปใ‹ใฎ RAM ๅฎน้‡ใงใฎ็ตๆžœใ‚’ใ€Community ใ‚ฟใƒ–ใ‹ Abyss ใฎ issue ใงๆ•™ใˆใฆใ‚‚ใ‚‰ใˆใ‚‹ใจๅŠฉใ‹ใ‚Šใพใ™ใ€‚

GPU 16 GB ๅ‘ใ‘ใฏ Qwen3.8-Flash-Next-IQ3_M-16G64G-GGUFใ€ GPU 12 GB + RAM 32 GB ๅ‘ใ‘ใฏ Qwen3.8-Flash-Next-E448-IQ2_M-12G32G-GGUF ใงใ™ใ€‚

ๅŒใ˜ๅคงใใ•ใฎ็‰ˆใจใฎๆฏ”่ผƒ

ใƒ€ใ‚ฆใƒณใƒญใƒผใƒ‰ ๅธธ้ง๏ผˆGPU + RAM๏ผ‰ KL๏ผˆ10 ๆœฌ๏ผ‰ 1 ไฝใฎไธ€่‡ด ๆ—ฅๆœฌ่ชžใฎ KL ใ‚ณใƒผใƒ‰ใฎ KL ็”Ÿๆˆ๏ผˆใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐ 3 ๅ•๏ผ‰
ใ“ใฎ็‰ˆ 93.7 GB 64.2 GB 0.068 90.3% 0.146 0.092 69.9 tok/s
unsloth UD-IQ4_XS 93.7 GB 64.9 GB 0.092 88.7% 0.195 0.116 71.1 tok/s

KL ใฏ NVFP4 ใฎใƒขใƒ‡ใƒซใจใฎๅทฎใงใ™๏ผˆ่ฉ•ไพก็”จใฎ 10 ๆœฌ ร— 1,023 ใƒˆใƒผใ‚ฏใƒณ๏ผ‰ใ€‚ใ€Œๅธธ้งใ€ใซใฏใ€ๅฑคใ”ใจใฎๅŸ‹ใ‚่พผใฟ่กจ๏ผˆ28.8 GBใ€1 ใƒˆใƒผใ‚ฏใƒณใซใคใๆ•ฐ่กŒใ ใ‘ SSD ใ‹ใ‚‰่ชญใ‚€๏ผ‰ใ‚’ๅ…ฅใ‚Œใฆใ„ใพใ›ใ‚“ใ€‚2 ใคใฎ็‰ˆใฏใƒ†ใƒณใ‚ฝใƒซใฎๅž‹ใŒๅŒใ˜ใงใ€้•ใ†ใฎใฏไธธใ‚ๆ–น๏ผˆLDLQใ€ๆ—ฅๆœฌ่ชžใจใƒใƒฃใƒƒใƒˆใฎใ‚„ใ‚Šใจใ‚Šใ‚’ๅ…ฅใ‚ŒใŸ่ชฟๆ•ด๏ผ‰ใ ใ‘ใงใ™ใ€‚

ๅฎŸ้š›ใซไฝฟใฃใŸใจใใฎ็ตๆžœ

Abyss ใฎใ‚ตใƒผใƒใƒผใ€ๆ€่€ƒใ‚ใ‚Šใ€seed ใ‚’ๅค‰ใˆใฆ 2 ๅ›žใ€ไธ‹ใฎ่จญๅฎšใงใ™ใ€‚

ใ“ใฎ็‰ˆ unsloth UD-IQ4_XS
ๆ—ฅๆœฌ่ชž 5 ๅ•๏ผˆใƒ“ใ‚ธใƒใ‚นใƒกใƒผใƒซใ€่งฃ่ชฌใ€่ฆ็ด„ใ€ๆ–‡ๅญฆใฎ็Ÿฅ่ญ˜ใ€่จˆ็ฎ—๏ผ‰ 10/10 ๅ›ž็ญ”ใ€ไบ‹ๅฎŸใจ่จˆ็ฎ—ใฏๆญฃใ—ใ„ 10/10
ใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐ 5 ๅ•๏ผˆ้š ใ—ใƒ†ใ‚นใƒˆ๏ผ‰ 5/5ใ€4/5 3/5ใ€5/5
8K ใƒˆใƒผใ‚ฏใƒณใฎๆ–‡็ซ ใ‹ใ‚‰ๆƒ…ๅ ฑใ‚’ๆŽขใ™ ๆˆๅŠŸ / ๆˆๅŠŸ ๆˆๅŠŸ / ๆˆๅŠŸ
็ฉบใฎๅ›ž็ญ” / ใƒ—ใƒญใƒณใƒ—ใƒˆใฎ็นฐใ‚Š่ฟ”ใ— ใชใ— ใชใ—
ๅ…ฅๅŠ›ใฎๅ‡ฆ็†๏ผˆ8K๏ผ‰ 1,680ใ€œ1,689 tok/s 1,670ใ€œ1,683 tok/s

้€Ÿๅบฆ

  • RTX 5090๏ผˆ32 GB๏ผ‰ใงใ€GPU 24 GB + RAM 64 GB ใฎ PC ใ‚’็œŸไผผใŸ่จญๅฎšใงใ™ใ€‚VRAM ใ‚’ 24 GB ใซๅˆถ้™๏ผˆSTRATA_GLM_VRAM_GB=24๏ผ‰ใ€RAM ใฎๆž  46 GB๏ผˆๅ…จไฝ“ 58 GB ใซๅˆถ้™๏ผ‰ใ€CPU 8 ใ‚นใƒฌใƒƒใƒ‰ใ€iGPU ใชใ—ใ€NVMe SSDใ€MTP ใƒ‰ใƒฉใƒ•ใ‚ฟใƒผใ‚ใ‚Šใ€‚
  • ใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐ 3 ๅ•๏ผš69.9 tok/s๏ผˆunsloth UD-IQ4_XS ใฏ 71.1๏ผ‰
  • ๆ—ฅๆœฌ่ชž 3 ๅ•๏ผš61ใ€œ77 tok/s
  • llama.cpp ใงใ“ใฎ็‰ˆใ‚’ --n-cpu-moe 34๏ผˆVRAM 24.2 GB๏ผ‰๏ผš26ใ€œ34 tok/s
  • Abyss ใฏ v1.3.108 ไปฅ้™ใŒๅฟ…่ฆใงใ™ใ€‚ใ“ใฎไธฆใณใฎ Q8_0 ใฎ down ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฏใ€ใใ‚Œใ‚ˆใ‚Šๅ‰ใฎ็‰ˆใงใฏๅ…ฅๅŠ›ใฎๅ‡ฆ็†ใง่ชญใฟ้ŽใŽใŒ่ตทใใฆใ„ใพใ—ใŸ๏ผˆunsloth ใฎ UD-IQ4_XS ใ‚‚ๅŒใ˜ใงใ™๏ผ‰ใ€‚

ไธญ่บซใƒปไฝœใ‚Šๆ–นใƒปๅ‹•ใ‹ใ—ๆ–น

  • 3 ใƒ•ใ‚กใ‚คใƒซใง 1 ใ‚ปใƒƒใƒˆใงใ™ใ€‚-00001-of-00003.gguf ใ‚’ๆŒ‡ๅฎšใ—ใฆใใ ใ•ใ„ใ€‚SHA256SUMS ใŒใ‚ใ‚Šใพใ™ใ€‚
  • ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใฏ gate/up ใŒ IQ3_S๏ผˆๅฑค 2 ใ ใ‘ IQ4_XS๏ผ‰ใ€down ใŒ IQ4_NL๏ผˆๅฑค 2ใ€4ใ€30ใ€46ใ€47 ใฏ Q8_0๏ผ‰ใงใ™ใ€‚ใใ‚Œไปฅๅค–๏ผˆใ‚ขใƒ†ใƒณใ‚ทใƒงใƒณใ€ใƒซใƒผใ‚ฟใƒผใ€ๅ…ฑๆœ‰ใ‚จใ‚ญใ‚นใƒ‘ใƒผใƒˆใ€ๅŸ‹ใ‚่พผใฟใ€ๅฑคใ”ใจใฎๅŸ‹ใ‚่พผใฟ่กจ๏ผ‰ใฏ unsloth UD-IQ4_XS ใฎใ‚‚ใฎใ‚’ใใฎใพใพไฝฟใฃใฆใ„ใพใ™ใ€‚
  • ไฝœใ‚Šๆ–นใฏไธŠใฎ่‹ฑ่ชžใฎ็ฏ€ใฎใจใŠใ‚Šใงใ™๏ผˆtrench-ggufใ€NVFP4 ใฎใƒขใƒ‡ใƒซใซใƒใƒฃใƒƒใƒˆใฎใ‚„ใ‚Šใจใ‚Š 181 ๅ•๏ผˆๆ—ฅๆœฌ่ชž 44 ๅ•๏ผ‰ใจๆ–‡็ซ ใ‚’ๆตใ—ใฆๆธฌใฃใŸ 126 ไธ‡ใƒˆใƒผใ‚ฏใƒณๅˆ†ใฎๆƒ…ๅ ฑใง LDLQ๏ผ‰ใ€‚
  • Abyss v1.3.108 ไปฅ้™ใงๅ‹•ไฝœใ‚’็ขบ่ชใ—ใฆใ„ใพใ™๏ผˆ64 GB ใฎ PC ใชใ‚‰ STRATA_GLM_RAM_GB=46๏ผ‰ใ€‚llama.cpp๏ผˆ2026-10-29 ใฎ ac8e1694d๏ผ‰ใงใ‚‚ใใฎใพใพๅ‹•ใใพใ™ใ€‚--n-cpu-moe 34 ใง VRAM 24.2 GB ใ ใฃใŸใฎใงใ€ๆœฌ็‰ฉใฎ 24 GB ใฎ GPU ใงใฏ 35 ไปฅไธŠใซใ—ใฆใใ ใ•ใ„ใ€‚ๅŒใ˜ๅฎŸ็”จใƒ†ใ‚นใƒˆใงใ€ๆ—ฅๆœฌ่ชž 5 ๅ•ใ™ในใฆใซๅ›ž็ญ”๏ผˆ็ฉบใฎๅ›ž็ญ”ใชใ—๏ผ‰ใ€ใ‚ณใƒผใƒ‡ใ‚ฃใƒณใ‚ฐ 3/5ใ€8K ใฎๆƒ…ๅ ฑๆŽขใ—ใ‚‚ๆˆๅŠŸใ—ใพใ—ใŸใ€‚

ใƒ•ใ‚ฃใƒผใƒ‰ใƒใƒƒใ‚ฏใฎใŠ้ก˜ใ„๏ผˆBETA ็‰ˆ๏ผ‰

  • ใŠไฝฟใ„ใฎ PC ใงใฎ้€Ÿใ•๏ผšAbyss ใง ./abyss.sh --bench --deep ใ‚’ๅฎŸ่กŒใ—ใ€abyss-bench.txt ใ‚’ ใ“ใกใ‚‰ใฎ issue ใƒ•ใ‚ฉใƒผใƒ ใซ่ฒผใฃใฆใใ ใ•ใ„๏ผˆใ‚ฟใ‚คใƒˆใƒซใฎ GPU ๅใ‚’ๆ›ธใๆ›ใˆใฆใใ ใ•ใ„๏ผ‰ใ€‚RAM ใฎๅฎน้‡ใจ SSD ใฎๆฉŸ็จฎใ‚‚ๆ›ธใ„ใฆใ‚‚ใ‚‰ใˆใ‚‹ใจๅŠฉใ‹ใ‚Šใพใ™ใ€‚
  • ๅฎŸ้š›ใซไฝฟใฃใŸใจใใฎๅ“่ณช๏ผš็ฉบใฎๅ›ž็ญ”ใ€่จ€่ชžใฎๅˆ‡ใ‚Šๆ›ฟใ‚ใ‚Šใ€็นฐใ‚Š่ฟ”ใ—ใ€unsloth ใฎ็‰ˆใ‚ˆใ‚Šๅผฑใ„ใจๆ„Ÿใ˜ใŸใจใ“ใ‚ใ€‚ๅ†็พใงใใ‚‹ใƒ—ใƒญใƒณใƒ—ใƒˆใŒใ‚ใ‚‹ใจไธ€็•ชๅŠฉใ‹ใ‚Šใพใ™ใ€‚
  • ใƒฉใ‚คใ‚ปใƒณใ‚นใฏๅ…ƒใฎใƒขใƒ‡ใƒซใจๅŒใ˜ใงใ™๏ผˆqwen-community-1.0๏ผ‰ใ€‚
Downloads last month
60
GGUF
Model size
177B params
Architecture
qwen4exp
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

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

Model tree for zephel01/Qwen3.8-Flash-Next-IQ3_M-24G64G-GGUF

Quantized
(601)
this model

Collection including zephel01/Qwen3.8-Flash-Next-IQ3_M-24G64G-GGUF