Darwin-35B-A3B-Mythos

Darwin-35B-A3B-Mythos

A measured Qwen3.6-35B-A3B reasoning derivative — frontier-class quality with published benchmarks, Korean capability, and a real on-device story. By VIDRAFT.

Most 35B-A3B derivatives on the hub ship a quant with no numbers. Darwin-35B-A3B-Mythos leads with measured results and stays fully reproducible under Apache-2.0.

Highlights (all measured, labeled)

• 34.7B total / ~3B active (A3B sparse MoE) — decode cost scales with the 3 active billion, so it is fast on small hardware.
• Korean-capable — no Korean-specialist model exists in the current trending set; Mythos speaks it.
• Apache-2.0, reproducible, honestly benchmarked.

Files

Coming: MTP-GGUF (native multi-token-prediction head, for self-speculative decode), NVFP4, FP8.

Run (llama.cpp)

Needs a recent llama.cpp build with qwen35moe support. Optimal on an 8GB card = experts on CPU, attention on GPU:

llama-cli -m mythos-35b-Q3_K_M.gguf -ngl 99 --n-cpu-moe 99 -c 8192 -p "..."

What's in v1 vs roadmap (honest)

v1 (this release): the measured Qwen3.6-35B-A3B derivative above — GPQA 86.4/70.7, Korean, full GGUF ladder, VKUE on-device numbers. Text-only (the vision tower is not included in this build).

Roadmap (not in v1 — will be labeled when shipped): DELPHI test-time reasoning layer (token-efficient thinking), native function-calling SFT, extended context, restored vision (mmproj), MTP-GGUF / NVFP4 format parity, and a live demo Space. We ship numbers when they are measured, not before.

Engines

• VKUE (ubiquity): the same weights run from a datacenter GPU down to an 8GB laptop — measured 20 tok/s on-device.
• VKAE (speed): VIDRAFT-optimized serving reaches 18,057 tok/s aggregate on a single B200 (measured).

Engine internals are proprietary; this card reports only measured results and standard open-tool run instructions.

Notes

• base_model: Qwen/Qwen3.6-35B-A3B — Darwin-35B-A3B-Mythos is a VIDRAFT derivative of that base.
• Benchmarks are labeled by method (greedy vs maj@8); do not compare across methods.
• Not multimodal in v1 (text-only). Do not use for tasks requiring vision until the vision build ships.