Two architectures for local agents

Two architectures for local agents An architecture diagram generated by Archify. Muse Spark · Meta teacher model · Meta Superintelligence Labs · 2026-08-10 · Apache 2.0 · larger, closed Muse Spark Meta teacher model larger, closed Logit distillation · pre + mid + post · Meta Superintelligence Labs · 2026-08-10 · Apache 2.0 · capability transfer Logit distillation pre + mid + post capability transfer Muse-Glimmer 30B · dense causal Transformer, 52 layers · Meta Superintelligence Labs · 2026-08-10 · Apache 2.0 · 29.6B, 131K ctx Muse-Glimmer 30B dense causal Transformer, 52 layers 29.6B, 131K ctx ViT-G/14 encoder · ~1.8B params, frozen · Meta Superintelligence Labs · 2026-08-10 · Apache 2.0 · separate vision ViT-G/14 encoder ~1.8B params, frozen separate vision DFlash drafter · 5 layers, block of 16 · Meta Superintelligence Labs · 2026-08-10 · Apache 2.0 · speculative DFlash drafter 5 layers, block of 16 speculative Qwen3.5 · architectural foundation · Alibaba Qwen team · 2026-08-14 · Apache 2.0 · lineage Qwen3.5 architectural foundation lineage Qwen3.8-27B · native vision-language, 64 layers · Alibaba Qwen team · 2026-08-14 · Apache 2.0 · 27B, 262K ctx Qwen3.8-27B native vision-language, 64 layers 27B, 262K ctx Gated DeltaNet · linear attention, 3 of 4 layers · Alibaba Qwen team · 2026-08-14 · Apache 2.0 · hybrid attention Gated DeltaNet linear attention, 3 of 4 layers hybrid attention MTP head · multi-token prediction · Alibaba Qwen team · 2026-08-14 · Apache 2.0 · speculative MTP head multi-token prediction speculative teacher outputs distilled vision tokens proposes 16-token blocks foundation attention layers decode head Meta Superintelligence Labs · 2026-08-10 · Apache 2.0 Alibaba Qwen team · 2026-08-14 · Apache 2.0 Legend Frontend Backend Database Cloud

Compressed

  • • Capability arrives from a larger teacher via logit distillation
  • • Textbook dense Transformer: every framework already knows it
  • • Vision is a separate frozen encoder; context 131K+

Grown in place

  • • Built on the Qwen3.5 hybrid: Gated DeltaNet + one Gated Attention per 4 layers
  • • Vision fused into the LM, incl. video; 262K native, 1M extensible
  • • MTP head trained in: speculative decoding is part of the model

The trade

  • • Distillation gives the smallest official recipe: 4-bit under 20GB, 24/32GB envelope, published tok/s
  • • Native growth gives a newer architecture and a higher ceiling, with the last mile left to the community (1,035 quantized builds on HF)