Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
Qwen3.8-Flash-Next Is Real Weights. Not a GX10 Sit.
Qwen/Qwen3.8-Flash-Next lastModified 2026-08-26T12:29 UTC. License qwen-community-1.0. 131 safetensor shards. Hugging Face usedStorage 360,013,002,208 bytes.
Key takeaways
- Qwen/Qwen3.8-Flash-Next lastModified 2026-08-26T12:29 UTC. License qwen-community-1.0. 131 safetensor shards. Hugging Face usedStorage 360,013,002,208 bytes.
- Card: 125B total / 6B active, plus 51B n-gram embedding and 4B MTP. Native 262K, YaRN to 1M. Image and video. They call this the architecture preview toward Qwen4.
- Unsloth GGUFs landed (lastModified 2026-08-26T15:54 UTC). I summed the shards: IQ1_S 72.5 GB, IQ1_M 74.5, Q2_K_XL 78.9, IQ3_XXS 82.0, Q3_K_XL 90.0, IQ4_XS 93.7, Q4_K_XL 111.3. Every quant still carries a ~50 GB shard. Q4_K_XL is too tight for 128GB unified. llama.cpp PR 27742 is open, not merged. Still not a sit.
- Community GGUF (not Unsloth): apetersson/Qwen3.8-Flash-Next-GGUF created 2026-09-07T13:25 UTC, lastModified 2026-09-08T12:16 UTC, usedStorage 334,704,069,344. I summed current LFS: Q5_K_M trunk 81,865,279,904; UD-IQ4_XS trunk 64,882,445,728; shared BF16 ngrams 102,400,491,712; shared MTP 2,786,568,256; mmproj 904,003,840. Trunk+ngrams is 167–184 GB on disk. GUIDE wants llama.cpp-shared-ngrams.patch (not stock). Still not a GX10 sit.
Local LLMs on NVIDIA Spark / ASUS GX10
Qwen posted Flash-Next this morning. The instrument is a Hugging Face card, last updated Aug 26 at 12:29 UTC. Qwen/Qwen3.8-Flash-Next. License name: qwen-community-1.0. 131 safetensor shards. I pulled the API: usedStorage is 360,013,002,208 bytes.
The card's own count: 125B parameters with 6B activated, plus 51B n-gram embedding and 4B MTP. Native context 262,144. They say YaRN extends it to 1,000,000. Image and video. They call this the experimental preview of the architecture that will underpin Qwen4. Qwen also says Qwen3.8-Flash is the official, more-production cut of this preview. That sentence is theirs.
I am not quoting their SWE table. I am not inventing tokens per second.
The fit test is still the file size. Official BF16 is about 360 GB of weights. Official FP8 (Qwen/Qwen3.8-Flash-Next-FP8) usedStorage is 185,553,536,918 bytes. RadixArk NVFP4 (RadixArk/Qwen3.8-Flash-Next-NVFP4) usedStorage is 135,242,371,978 bytes. None of those official or NVFP4 weights fit 128GB unified.
Unsloth posted GGUFs this afternoon. The card is unsloth/Qwen3.8-Flash-Next-GGUF, last updated Aug 26 at 15:54 UTC. I summed the .gguf shards on the tree:
- UD-IQ1_S: 72,546,461,344 bytes (72.5 GB)
- UD-IQ1_M: 74,538,755,776 (74.5 GB)
- UD-Q2_K_XL: 78,869,128,864 (78.9 GB)
- UD-IQ3_XXS: 81,961,823,936 (82.0 GB)
- UD-Q3_K_XL: 89,986,353,824 (90.0 GB)
- UD-IQ4_XS: 93,682,584,224 (93.7 GB)
- UD-Q4_K_XL: 111,334,654,784 (111.3 GB)
Every quant still carries a ~50 GB shard. That is the n-gram / PLE table riding along. Q4_K_XL is too tight for 128GB unified even before KV. A 72.5 GB file list is not a sit.
To run, Unsloth's own card points at llama.cpp PR 27742 (open, not merged) or Unsloth Desktop. I am not checking out that branch. I am not inventing tok/s.
A community package showed up later under apetersson/Qwen3.8-Flash-Next-GGUF. Not Unsloth. Card created Sep 7 at 13:25 UTC; lastModified Sep 8 at 12:16 UTC. API usedStorage is 334,704,069,344 bytes. Layout is two standalone trunks plus a shared directory: Q5_K_M-BF16-Ngrams and UD-IQ4_XS-BF16-Ngrams, with shared/ngrams-Qwen3.8-Flash-Next-BF16.gguf, shared MTP, and mmproj. I summed the current LFS tips myself:
- Q5_K_M trunk: 81,865,279,904 bytes (76.24 GiB)
- UD-IQ4_XS trunk: 64,882,445,728 (60.43 GiB)
- shared BF16 ngrams: 102,400,491,712 (95.37 GiB)
- shared MTP Q8_0: 2,786,568,256 (2.60 GiB)
- mmproj f16: 904,003,840 (0.84 GiB)
Each of those files alone is under ~110 GB. The runnable set is trunk plus the shared BF16 table: IQ4+ngrams 167,282,937,440 bytes; Q5+ngrams 184,265,771,616. That is still over 128GB on disk before KV. The card says the recipes are aimed at 128GB systems with --lazy-mode on, and their smoke tests are on an Apple M1 Ultra with Metal — peak wired numbers on that card are theirs, not mine. Stock llama.cpp does not resolve qwen4exp.ngram_file. Their GUIDE points at commit 465e49b9 plus llama.cpp-shared-ngrams.patch (that patch subsumes the older MTP patch; do not stack both). I did not apply the patch. I did not load this on the GX10. I am not inventing tok/s. Still not a sit.
I run Grok when it earns it, and a GX10 when I want the weights in the room. This family is still on the far side of that split. The 27B receipts stay the 27B receipts. If a later quant actually sits in 128GB and I load it, that note comes next. Not this card, and not a download I did not start.
Frequently asked questions
Yes. The Hugging Face card Qwen/Qwen3.8-Flash-Next was last updated Aug 26, 2026 at 12:29 UTC. Qwen's blog is qwen.ai/blog?id=qwen3.8-flash-next. License name on the card is qwen-community-1.0.
No. Official BF16 is about 360 GB. Official FP8 is about 186 GB. RadixArk NVFP4 usedStorage is 135,242,371,978 bytes. Unsloth GGUFs are real files now: IQ1_S sums to 72,546,461,344 bytes, Q4_K_XL to 111,334,654,784. Every quant still carries a ~50 GB shard. File size is not a sit. I did not load any of these on the GX10. Community trunks under apetersson/Qwen3.8-Flash-Next-GGUF alone look under ~110 GB, but trunk plus the shared BF16 ngrams table is 167–184 GB on disk; the lazy-mode claims are that card's M1 Ultra story, not a GX10 load.
No. The 27B receipts stay the 27B receipts. This is a different family: 125B / 6B active plus a 51B n-gram table. I am not rewriting that sit.
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.