Ornith 1.5 35B
A registered and benchmarked hybrid MoE model. Guild’s README classifies the Qwen3.5 MoE / Ornith execution path as experimental and under active development.
- Architecture
- qwen35moe
- Parameter count
- 35B nominal model name; upstream count not independently verified
- Active parameters
- Not verified; 8 of 256 routed experts per layer
- Formats
- GGUF input · GPACK container · optional GEXEC cache
- Registry quantization
- Q4_K_M
- Registry size on disk
- 20.22 GiB (21,713,463,040 bytes)
- Layers / routed experts
- 40 layers · 8 of 256 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
The upstream repository named in the registry could not be found on 2026-10-22. Use an existing lawful local model artifact; download availability is unverified.
Phase 6, 4,096 populated prompt / 32,768 capacity, host-only FP16 KV: 2,662 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
- Registry quantization: Q4_K_M. Phase 6/6.5/6.6 reports include actual model runs.
- The benchmark artifact has no MTP weights; lookup and session-local retrieval do not need a draft model.
- Although registry metadata says 262,144 context tokens, the documented dense-attention limit is 32,768 total in the Phase 6 benchmark configuration.
- No current upstream download is presented as verified.