AI Infra Interviews logo

ERNIE / HCA / MLA-DSA latent MoE

ERNIE-5.5-Flash-Base

Start with the training stage and license

Base-weight releaseSources checked 2026-11-16

Useful for teams evaluating continued training or post-training on a new base. It is not a drop-in chat replacement simply because a completion server starts.

280Breported total parameters
14Bactive parameters per token
1,048,576configured token positions

Publisher count excludes 0.36B additional ECHO parameters. Active count includes the output head, not embedding lookup. Layer counts describe the 43-layer backbone, excluding its prediction layer.

Evaluate it for

Adaptation research with permitted use and a compatible model-specific runtime.

Choose another path when

The requirement is a ready-to-use conversational agent or the intended territory or commercial terms are incompatible.

Release context

Why this release matters

Thirty-seven hierarchical compressed-attention layers group history at a 128-token ratio; six MLA-DSA layers use a different retained representation. The whole cache does not shrink by 128 times.

The applicable model license identifies November 5 as its release date; the repository existed earlier.

Release evidence · Follow the ERNIE timeline →

Deployment starting points

Which hardware, how many, at what precision?

Start with an exact artifact and an engine that supports it. A configuration below is evidence of a documented path; its device count does not promise a particular throughput or concurrency.

No certified hardware configuration in this edition

The requirement is a ready-to-use conversational agent or the intended territory or commercial terms are incompatible.

The pinned model card establishes the architecture. It does not, by itself, prove that a particular GPU count runs this checkpoint at your target context. The capacity screen below helps rule out undersized allocations; a successful load and workload test are still required.

288.42 GB of tensor data in the pinned index

This is the index’s reported tensor payload, in decimal GB. It excludes file headers and runtime memory. Mixed precision, conversion and host/device placement determine how much GPU memory the loaded model needs.

Inspect the exact index metadata →
Capacity screen: would the weights fit?

This estimates inference memory for a hypothetical uniform precision. It is useful for rejecting an allocation that is too small. It does not establish a working deployment or estimate training memory.

A uniform-precision GPU count would hide this model’s component layout. Publisher count excludes 0.36B additional ECHO parameters. Active count includes the output head, not embedding lookup. Layer counts describe the 43-layer backbone, excluding its prediction layer. Use the documented artifact and placement path above; no GPU count is inferred here.

Architecture in practice

What the serving engine has to do

Compression applies to a particular state representation.

  1. HCA layer37 backbone layers
    Compressed history128-token grouping
  2. MLA-DSA layer6 backbone layers
    Latent historyselection and retained state
HCA layer37 backbone layersCompressed history128-token groupingMLA-DSA layer6 backbone layersLatent historyselection and retainedstate

Sparse reads do not imply proportional storage savings.

Sum both layer families before adding runtime overhead.

ERNIE 5.5 Flash Base backbone schematic. Thirty-seven HCA layers compress groups of token history; six MLA-DSA layers use latent history and sparse selection. ECHO and runtime buffers add separate state. This is not a literal execution ordering.

Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.

Inspect the full checkpoint specifications
Inspected fields for this exact revision. KV heads alone do not describe latent, hybrid or shared-cache layouts.
Exact checkpointbaidu/ERNIE-5.5-Flash-Base
Text model typeernie5_5_moe
Layers43
Hidden width4,096
Attention / KV heads64 / 1
Head dimensionNot verified
Routed / selected experts512 / 10
Configured positions1,048,576
Documented extensionNot recorded for this checkpoint
Inputs → outputtext → text
License metadataERNIE 5.5 Flash Base License v1.0 · custom terms

37 HCA layers; 6 MLA DSA layers. These counts describe the configured pattern.

Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.

A useful comparison

What else belongs on the shortlist?

The alternatives are post-trained serving candidates. Comparing their chat scores directly with this base would conflate architecture and training stage.

17 Oct 2026

DeepSeek-V4.1-Flash-1017

Same footprint, new weights: a clean post-training comparison

552B reported total · text + image

27 Oct 2026

Ling-3.1-flash-1019

A post-training update needs its own deployment evidence

560B reported total · text

These are editorial comparison candidates. We have not run a matched quality or serving benchmark, so this is not a ranking.

Prepare to explain it

The interview lesson

Separate 131,072-token training context from the released 1,048,576-position configuration. Neither establishes tested production capacity.

Attention layouts · Quantization · Courses and worked examples

For a deployment evaluation, record exact weights, precision, engine, device count, interconnect, prompt/output lengths and concurrency. Report task success, errors, TTFT, TPOT and useful throughput together.

Sources and citation

Our assessment is editorial judgment based on the linked architecture and deployment evidence. Model facts are publisher-reported or attributed to runtime maintainers; no independent GPU benchmark was run.

AI Infra Interviews, “ERNIE-5.5-Flash-Base: release, architecture and deployment”, checked 2026-11-16. Preserve this date and the exact checkpoint when citing.

Model and research notes as JSON · Baidu profile