Qwen
Alibaba Qwen
One family spans tiny dense models, coding MoEs and conditional memory.
Muse Glimmer / Lab profile
Meta
A local-agent release includes the runtime around the model.
First time comparing models? Learn the terms before the architecture →
This reference currently covers Muse Glimmer, not the full historical Llama catalog or hosted Muse products. Glimmer is useful for understanding a local-agent release as a package: a dense multimodal model, quantized artifacts, a speculative drafter and device-specific evaluations.
Architecture trajectory
What connects these models?
The 29.6B model combines local and full attention with a small configured KV-head count. DFlash adds a drafting-and-verification path. A speed comparison therefore needs to state whether the drafter is enabled, its precision and its acceptance behavior.
Fast drafts help only when verification accepts them.
Acceptance rate and verification cost decide the gain.
Count drafter memory alongside the target model.
Newest first / grouped sequence, not a time scale
The Muse Glimmer timeline
Muse-Glimmer-30B
A local agent needs more than a small download
The 29.6B dense model uses alternating local and full attention with only two configured KV heads. Meta’s published local-device work makes quantization and speculative decoding part of the release story.
The publisher card gives August 2026, without a day. It is grouped after dated August releases; that placement does not imply it was older.
Which other labs should you compare?
Qwen3.8-27B and Gemma 4 31B are useful dense multimodal alternatives. Compare complete local-agent tasks: correct tool actions, recovery from failure, memory at the target context and wall-clock time. More generated tokens per second may not mean more completed work.
Gemma
Google DeepMind
Gemma is a family of different design choices, not just different sizes.
Interview preparation
What you should be able to explain
Explain speculative decoding with an acceptance loop, then identify how quantized main weights, drafter weights and request cache contribute separately to memory.
Study the underlying systems → · Get the illustrated report →
Sources and scope
Checked 13 September 2026. Assessments and comparison candidates are our interpretation of the cited designs, not measured cross-lab rankings. Each timeline entry links to its release evidence; model pages preserve pinned card and configuration sources.