Qwen3.8-Flash-Next
Guild’s primary development target. Hybrid recurrent and attention execution with routed experts, per-layer embeddings, and an optional MTP draft path.
- Architecture
- qwen4exp
- Parameter count
- 176.94B stored parameters (Hub GGUF metadata)
- Active parameters
- Not verified in the available metadata
- Formats
- GGUF input · GPACK container · optional GEXEC cache
- Registry quantization
- UD-IQ4_XS
- Registry size on disk
- 87.25 GiB (93,682,584,224 bytes)
- Layers / routed experts
- 48 layers · 10 of 512 experts active per layer
- Context metadata
- 262,144 tokens — see runtime limits below
Guild snapshot 140f6f7. Unknown fields are left unverified rather than estimated.
Availability
Public upstream repository verified on 2026-10-22. Guild registry selects UD-IQ4_XS; other Hub quantizations are not automatically validated in Guild.
Verified upstream revision ↗
guild pull qwen3.8-flash-next
Phase 6, 4,096 populated prompt / 32,768 capacity, host-only FP16 KV: 7,310 MiB sampled VRAM on GTX 1070. This is an observed configuration, not a minimum requirement.
Measured host: 2× Xeon Gold 6130, 376 GiB RAM, GTX 1070 8 GiB. System-RAM requirements are not established by sampled VRAM. Context, KV, and workspace choices matter. Read hardware guidance ↗
Compatibility notes
- Hub GGUF parameter totals describe stored tensors, not per-token active computation. Do not equate them to marketing model sizes.
- Guild registry geometry: 48 layers, 512 routed experts, 10 selected per layer.
- Registry context metadata is 262,144 tokens; runtime limits and usable context depend on the runner, memory, and configuration.
- The Hub lists multimodal assets; this catalog does not claim Guild vision or video support.