AI Infra Interviews logo

Ling / Lab profile

InclusionAI

Hybrid state and checkpoint identity decide the rollout.

The covered Ling releases share a 560B expert store and a 25B active path. Their relevance is the combination of recurrent delta attention and gated latent attention, followed by a separately evaluated post-training update.

Architecture trajectory

What connects these models?

Count 63 KDA and nine gated MLA layers separately. The nine history-retaining layers prevent a whole-model constant-state assumption.

Two kinds of layers retain two kinds of state.

A token enters a layerthe layer type selects itspathRecurrent layerupdate a compact stateAttention layerread retained tokenhistoryRecurrent statemaintained for thissequenceAttention cachelayout depends on themodel

Hybrid attention does not mean zero request state.

A full-attention formula cannot describe every layer.

Schematic of a hybrid attention stack, not a literal layer ordering. Recurrent layers update a compact state; attention layers retain their own history. Both kinds of state matter for prefix reuse and concurrent requests.

Newest first / grouped sequence, not a time scale

The Ling timeline

  1. Ling-3.1-flash-1019

    A post-training update needs its own deployment evidence

    The two releases share 560B total and 25B active counts. Their artifact identity, task results and rollout acceptance still need separate records.

    The repository records the weight upload on this date. The 1019 suffix is not proof of an October 19 public weight release.

    Release evidence ↗

  2. Ling-3.1-flash

    Count recurrent and latent state separately

    Sixty-three Kimi Delta Attention layers retain recurrent state; nine gated multi-head latent-attention layers retain history. Cache reuse must restore both.

    Dated weight-upload commit; not inferred from repository creation or an API listing.

    Release evidence ↗

Which other labs should you compare?

Use MiniMax as another hybrid cluster candidate and a smaller Qwen checkpoint when residency is the limiting constraint.

MiniMax M3

MiniMax

Study how sparse attention changes what the GPU reads.

Qwen

Alibaba Qwen

One family spans tiny dense models, coding MoEs and conditional memory.

Interview preparation

What you should be able to explain

Reconcile the card’s million-token claim with its 262,144-position pinned config before approving a long-context service.

Study the underlying systems → · Get the illustrated report →

Sources and scope

Covered checkpoints carry their own research dates. 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.