Modelpublic

6block/Qwen3.8-Flash-Next-GGUF

sourceHugging Faceotherupdated 24d agoView on Hugging Face

0likes1.1kdownloads

Model Card

Qwen3.8-Flash-Next GGUF

GGUF quantizations of Qwen/Qwen3.8-Flash-Next, made with llama.cpp.

180B total parameters. A MoE core of 125B (48 layers, 512 experts, top-10 plus one shared expert) plus a 51B PLE n-gram lookup table. Hybrid attention: Gated DeltaNet linear attention interleaved with Qwen Sparse Attention (QSA) blocks, wrapped in hyper-connections. Context length up to 262K. The one MTP draft layer in the checkpoint is excluded at conversion time.

Quantized from the official BF16 weights. Every tier is imatrix-calibrated and ships as 8 shards.

Quantizations

Quant	Size	Shards	BPW	PPL (wikitext-2)
BF16 master	329.7 GiB	8	15.73	1.7164 ± 0.04795
`IQ4_XS`	118.8 GiB	8	5.67	1.8378 ± 0.05554
`IQ3_XXS`	107.4 GiB	8	5.13	2.0528 ± 0.06396
`IQ2_XS-layered`	83.5 GiB	8	3.98	2.7833 ± 0.10058
`IQ1_M-layered`	76.4 GiB	8	3.65	4.4105 ± 0.19322

The master row is not a file in this repo. It is listed so the numbers above have a reference point — the BF16 master fits on the PPL machine, so tiers are compared against a real baseline.

The two low tiers use layered quantization (IQ2_XS-layered, IQ1_M-layered): the 51B PLE n-gram table tolerates heavy compression (measured: Q8_0 → Q4_0 costs nothing), so it is packed at Q4_0, and the space saved is spent on the expert gate/up projections instead — which improves PPL by ~14% at a smaller file size than the uniform tier. High tiers stay uniform since they already sit close to the master.

What is protected

The layers that would hurt most under low-bit compression are kept at higher precision:

Tensors	Type	Reason
`ffn_gate_inp` / `ffn_gate_inp_shexp`	F32	MoE router; compressing it routes to the wrong experts
`ffn_*_shexp`	Q8_0	the shared expert runs on every token
`ssm_a` / `ssm_conv1d` / `ssm_dt` / `ssm_alpha` / `ssm_beta` / `ssm_norm` / `ssm_out`	F32	Gated DeltaNet linear-attention state; low bit-width destroys long-range recall
`hc_attn_` / `hc_ffn_` / `output_hc_*`	F32	hyper-connection streams, every layer
`ple_key` / `ple_value` / `ple_conv1d`	Q8_0	PLE projection layers
`per_layer_token_embd` (51B n-gram table)	Q4_0	n-gram counts, highly compression-tolerant
`token_embd` / `output`	Q6_K	a global type would otherwise squeeze these hard

With 512 experts the expert layers dominate the file, so protecting everything else is cheap.

Usage

bash

# Point at the first shard; llama.cpp finds the rest on its own.
llama-cli -m Qwen3.8-Flash-Next-IQ4_XS-00001-of-00008.gguf

Do not pass -ngl manually — llama.cpp fits layers to free VRAM by itself, and any explicit -ngl aborts that fitting. Leave it off.

Every tier ships as 8 shards. Download all 8 into one directory — you only ever name -00001-of-00008 on the command line.

About the PPL numbers

wikitext-2 test, n_ctx=512, 12 chunks, every tier through the exact same command. These numbers are only comparable within this table. Do not compare them against PPL figures published by other repos.

The BF16 master row is the measured baseline. IQ2_XS-layered / IQ1_M-layered use layered quantization (see above) — at the same size they beat the uniform tiers by a wide margin.

License

Qwen Community License 1.0, inherited from Qwen/Qwen3.8-Flash-Next — this is not MIT. Quantized by 6block.