Qwen3.8-Flash-Next for DACAN

Weights of Qwen3.8-Flash-Next packed for DACAN — our engine, a fork of Strata rebuilt for two RTX 2080 Ti (22 GB, NVLink) and two AVX-512 Xeon sockets. DACAN runs this one model family only.

Status (03.10.2026): complete, 27 files (260.6 GB). The dense GGUF was checked against the file the DACAN service runs: three greedy answers (1 984, 1 931 and 596 tokens) came out identical character for character.

The model and its architecture

Qwen4ExpForConditionalGeneration, model_type: qwen4_exp, GGUF general.architecture: qwen4exp:

• 125B parameters with 6B activated per token, plus a 51B n-gram embedding table and a 4B MTP layer
• 48 layers: 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Qwen Sparse Attention → MoE)), hidden size 2560
• Gated DeltaNet: 48 V heads, 16 QK heads; Qwen Sparse Attention: 24 Q heads, 2 KV heads
• MoE: 512 experts, 10 routed + 1 shared per token, expert width 640
• Gated residual (widened residual streams), n-gram embedding at layer 2, 1 MTP layer
• Context 262,144 tokens natively

DACAN checks this geometry on load (48 layers, 2560, 512 experts, 10 active, 24 / 2 heads) and refuses anything else, so fine-tunes of Qwen3.8-Flash-Next with the same shape work, other models and pruned variants (REAP etc.) do not.

Two layouts: NVFP4 + Q8 (DACAN) and Q8_0

NVFP4 + Q8 (DACAN) is our name for the layout the DACAN service runs: NVIDIA's NVFP4 routed experts, everything else that is large in Q8_0, the small sensitive matrices left in BF16. The experts are about 95 % of the weights and are read from RAM on every token, so their size sets the speed; the parts every token passes through stay at 8 bits.

The Q8_0 layout swaps the experts for ours in Q8_0 (128 GB) and keeps the rest; it needs 120 GiB of RAM for the experts instead of 63 GiB and answers slower (on Swift 1.5: 49 tok/s against 58, median of 9 runs). The same NVFP4 + Q8 (DACAN) mix of the fine-tune Swift 1.5, with experts we quantized from its BF16 weights, is in tirex2001/Swift-1.5-Qwen3.8-Flash-Next-DACAN.

Files

Layout. Put the dense GGUF and the *-experts.bin you use in one folder: pack-*/native_experts.txt names the experts file without a path and DACAN looks for it next to the --native GGUF. The service config we run (NVFP4 + Q8 (DACAN)) is --pack pack-nvfp4 --native Qwen3.8-Flash-Next-dense-Q8_0.gguf --ple-gguf Qwen3.8-Flash-Next-PLE-FP8-Q8_0.gguf --mtp mtp --expert-profile data/profile_other_2609.bin --second-card 1 --second-card-usage data/usage_other_2609.bin --max-context 262144 --kv int8 --spec 4 (full list in the DACAN README).

How to run: see the DACAN README. Measured 03.10.2026 on 2× RTX 2080 Ti + 2× Xeon 8368 (NVFP4 + Q8 (DACAN), free node, greedy):

The answer speed follows the MTP drafts: 85–89 % accepted on code and lists, 57 % on English prose, 32 % on Russian. Context 256K (a needle found up to 195K). Details: README-2x2080Ti.

Licenses

The model: Qwen Community License 1.0, copyright (c) 2026 Qwen. The NVFP4 experts are derived from NVIDIA's checkpoint and licensed by NVIDIA Corporation under the NVIDIA Open Model License.

Support the project · Поддержать проект

Everything here — the quants, the engine, the measurements — is made and published for free. If it helped you run a model on your own hardware, you can say thanks with a donation. Если это помогло вам запустить модель на своём железе, можно поблагодарить донатом.

⚠️ Send only Tron-network assets (USDT TRC-20, TRX) to the Tron address and only Ethereum-network assets to the Ethereum address; anything sent over the wrong network is lost. Сеть важна: отправленное не в той сети пропадёт.

Thank you! Спасибо! 🙏