hfviewer by embedl

Interactive model architecture

Architecture graph for embedl/Qwen3.5-0.8B-FlashHead.

Interactive architecture graph for embedl/Qwen3.5-0.8B-FlashHead, visualized from Hugging Face model metadata.

Paste a Hugging Face link to visualize it

No export step, no config hunt, no model surgery. Paste the link and inspect the graph.

Graph structure Understand the high-level graph structure of different transformer models.

URL magic You can replace huggingface.co with hfviewer.com in the url to view it.

Chrome Extension! View each model directly on Hugging Face! Install extension

Embed in model card Embed the architecture graph directly in your Hugging Face model card. Add to your model card!

Quote block direction

Seventy quote block variants.

Same article quote, seventy polished hfviewer-style treatments. These are intentionally isolated from the real article styling so they can be compared side by side.

Variant 01

Prism rail

Closest to the current article language, but with a more deliberate glow and depth.

Variant 02

Folded glass

A subtle document-card feel with a bright folded corner instead of a classic quote mark.

Variant 03

Signal frame

A technical conic border and faint scan-lines, closer to model-graph instrumentation.

Variant 04

Editorial shell

The older left-accent idea, but as a gradient shell behind an inset article card.

Variant 05

Graph glow

Node-like dots and soft graph geometry integrated into the quote surface.

Variant 06

Layer stack

Feels like a compact stack of model-card layers, with enough weight for lead quotes.

Variant 07

Insight capsule

A more iconic pill shape with a glowing green anchor, good for short high-signal quotes.

Variant 08

Aurora card

A richer atmospheric quote card using the hfviewer blue/green/pink glow system.

Variant 09

Blueprint

A crisp technical panel with subtle grid structure and a low bottom accent line.

Variant 10

Minimal pro

The quietest option: premium, readable, and less decorative while still themed.

Variant 11

Margin rail

An open editorial quote: no full card, just a luminous margin cue and a soft reading surface.

Variant 12

Corner field

Open corners instead of a container, like the article is selecting an insight on the canvas.

Variant 13

Underline pulse

Mostly typography: a strong quote line with a layered hfviewer underline rather than a block.

Variant 14

Diagonal shard

A more angular model-paper surface, deliberately avoiding the normal rounded rectangle.

Variant 15

Terminal line

A technical prompt-style quote for article passages that should feel precise and executable.

Variant 16

Node path

The quote starts at small graph nodes, making the article-to-visualizer relationship feel native.

Variant 17

Manuscript mark

A marked-up research-note treatment with line rhythm and small colored anchor points.

Variant 18

Theorem strip

A compact paper-like statement row; structured, open, and less visually boxed.

Variant 19

Waveform

A subtle signal/animation-inspired quote treatment with no explicit card boundary.

Variant 20

Glass cutout

A cut-corner glass panel: still premium and readable, but less like a conventional roundrect.

Variant 21

Dot anchor

A single characteristic glowing dot becomes the quote’s origin point and vertical guide.

Variant 22

Orbit dot

The dot sits inside a small orbit, making the quote feel like an inspected graph state.

Variant 23

Node chain

A small row of glowing graph nodes leads into the quote without needing a traditional box.

Variant 24

Pinned insight

A glowing dot pins the top-left corner, giving the quote a precise annotated-paper feel.

Variant 25

Constellation

Several small glowing dots and faint connections give the quote a more model-graph-native identity.

Variant 26

Beacon

The glowing dot acts like a signal beacon, with soft rings and a directional trail.

Variant 27

Split rail dot

A dot interrupts a vertical rail, so the quote feels selected without becoming a boxed card.

Variant 28

Endpoint dots

The quote is suspended between two glowing endpoints, like an edge label in the graph.

Variant 29

Spotlight dot

A centered dot shines down onto the quote, making it feel like a highlighted article thesis.

Variant 30

Inline dot badge

The glowing dot becomes part of the text rhythm itself rather than a surrounding container.

Variant 31

Dot anchor clean

Variant 21 without the vertical green line: just a dot, a soft field, and strong text.

Variant 32

Dot halo

The dot gets a thin halo ring instead of a line, giving it more presence without boxing the quote.

Variant 33

Left glow wash

The dot is supported by a broad atmospheric glow rather than any visible edge or rail.

Variant 34

Dot notch

A tiny horizontal trace replaces the vertical line, keeping the quote anchored but cleaner.

Variant 35

Inline leader dot

The dot becomes the first typographic unit, making the quote read like a premium thesis line.

Variant 36

Floating pin

The dot floats above the quote like a selected graph node, without needing a surrounding block.

Variant 37

Soft dot field

The closest to Variant 21, but with only the dot and background atmosphere; no line or trace.

Variant 38

Blue dot anchor

A blue version using the other hfviewer signature color, lighter and more graph-view aligned.

Variant 39

Dot with echo

The main green dot gets a small pink echo dot, creating more visual rhythm without a rail.

Variant 40

Thesis mark

The most minimal Variant 21 derivative: a small glowing dot as a quiet thesis marker.

Variant 41

Thesis soft glow

Variant 40’s quiet dot, but with Variant 37’s soft atmospheric glow and a subtle breathing pulse.

Variant 42

Wide soft field

A slightly larger soft field around the thesis dot, still no rail and no obvious quote container.

Variant 43

Small breathing mark

A smaller, calmer dot with the same soft-field logic; more understated than Variant 41.

Variant 44

Left bloom

The dot breathes from inside a wider left-side bloom, giving the quote more presence without a border.

Variant 45

Quiet breath

The least decorative animated version: only a faint field and a slow breathing dot.

Variant 46

Blue-green field

The green dot stays central, while the field borrows more of the hfviewer graph-view blue.

Variant 47

Edge thesis dot

The dot sits just outside the text rhythm, closer to a margin mark than a quote block.

Variant 48

Double breath field

A second faint blue echo breathes behind the dot, adding depth without adding another visible marker.

Variant 49

Top note

The glowing dot moves above the quote, like a quiet article annotation rather than a left rail.

Variant 50

Underline glow thesis

The minimal dot stays, but the breathing glow also runs into a very subtle bottom underline.

Variant 51

Unclipped thesis soft glow

Variant 41 with the same breathing atmosphere, but the left glow is allowed to continue outside the comparison card.

Variant 52

Unclipped wide soft field

Variant 42 with a wider visible bloom that bleeds beyond the quote’s left edge instead of being cut flat.

Variant 53

Unclipped small breathing mark

Variant 43 with the same small mark, now letting the soft halo extend naturally past the sample boundary.

Variant 54

Unclipped left bloom

Variant 44 with the strongest left-side glow, designed specifically to read as a seamless bloom outside the quote.

Variant 55

Unclipped quiet glow

Variant 45 with a lower-intensity bleed, useful if the glow should feel present but not decorative.

Variant 56

Unclipped blue-green glow

Variant 46 with the blue secondary field kept visible while the green thesis glow bleeds past the left edge.

Variant 57

Unclipped edge spark

Variant 47 keeps the dot close to the edge, but removes the hard crop so it feels intentionally anchored there.

Variant 58

Unclipped double breath

Variant 48 with both green and blue breathing fields visible beyond the quote, giving the mark more depth.

Variant 59

Unclipped top note

Variant 49 with extra room for the upper note glow, preventing the top-left atmosphere from being clipped.

Variant 60

Unclipped underline glow

Variant 50 with the underline and left glow free to breathe past the sample boundary instead of ending abruptly.

Variant 61

Static thesis glow, card-clipped

Variant 51 without breathing motion: the blob can leave the quote, but the article card remains the outer mask.

Variant 62

Static wide field, card-clipped

Variant 52 made still, with the glow contained by the article border instead of the quote boundary.

Variant 63

Static small mark, card-clipped

Variant 53 with no motion; a quieter dot that still casts past the quote line without escaping the card.

Variant 64

Static left bloom, card-clipped

Variant 54 without animation: the strongest bloom continues beyond the quote, then resolves at the article border.

Variant 65

Static quiet glow, card-clipped

Variant 55 with the calm field locked in place, useful for a less animated article quote treatment.

Variant 66

Static blue-green glow, card-clipped

Variant 56 without motion, preserving the blue secondary glow while clipping only at the article frame.

Variant 67

Static edge spark, card-clipped

Variant 57 made still, so the edge dot feels anchored while its glow is masked by the card rather than the quote.

Variant 68

Static double glow, card-clipped

Variant 58 with fixed green and blue glow masses, contained by the article card instead of the quote rectangle.

Variant 69

Static top note, card-clipped

Variant 59 without motion; the top-note glow has room to leave the quote while staying inside the article card.

Variant 70

Static underline glow, card-clipped

Variant 60 made still, keeping the underline haze clipped by the article card instead of the quote itself.

Editor - interactive article

Intro

Model Graph preview

Node

What detail would help?

Screenshots

Optional: draw a rectangle around the part that needs more detail.

Popular models

Click a model to open it.

MODEL PAGES WITH HFVIEWER

Community showcase.

Hugging Face authors are adding the hfviewer model card directly to their READMEs.

Add to your model card!

AxiomicLabs/GPT-X2-125M

A 125M custom GPT-X2 language model with RoPE, SwiGLU, grouped-query attention, and curriculum-trained code/math normalization.

Open visualization View on HF

AxiomicLabs/GPT-S-5M

A compact 5M GPT-S language model using RoPE, SwiGLU, grouped-query attention, and exclusive shared attention for small-model reasoning experiments.

Open visualization View on HF

kalyan-ks/ettin-68m-nemotron-pii

A MIT-licensed Ettin encoder fine-tuned for PII token classification on Nemotron-PII, reporting 96.27 F1 on the test split.

Open visualization View on HF

BEE-spoke-data/mega-ar-126m-4k

A compact 126M language model built with MEGA rather than a standard transformer stack, with a 4096-token context length.

Open visualization View on HF

DominicTWHV/Horizon-1-Text-Large

A larger, more modern Constellation-One variant for Cockatoo, fine-tuned from answerdotai/ModernBERT-large.

Open visualization View on HF

BEE-spoke-data/pegasus-x-base-synthsumm_open-16k

A Pegasus-X based long-document summarization model fine-tuned on synthetic summaries for general summarization with long contexts.

Open visualization View on HF

juiceb0xc0de/bella-bartender-gemma-e2b

A Gemma 4 E2B based conversational model tuned toward a more direct, personal voice rather than a generic assistant cadence.

Open visualization View on HF

databoyface/bert-base-uncased-ome-v5

A BERT based OME v5 classifier for English emotion examples, fine-tuned over 26 categories and reporting 98.63 percent accuracy on eval.

Open visualization View on HF

databoyface/distilroberta-base-ome-v5.2

A DistilRoBERTa-based OME v5.2 emotion classifier for English text, fine-tuned over 26 categories and reporting 98.03 percent eval accuracy.

Open visualization View on HF

AINovice2005/ModernBERT-base-lora-cicflow-1m-r4

A LoRA fine-tune of ModernBERT-base for binary classification, designed for high recall with controlled false positives.

Open visualization View on HF

If you are interested in deploying these models to edge devices, check out our other products:

embedl deploy Quantization solved embedl hub Compliant MLOps embedl models optimized genai

Release note

Introducing hfviewer

The Hugging Face ecosystem already has model cards, Spaces, checkpoints, benchmarks, and demos. What it has still been missing is a fast general-purpose way to see how a model is put together. We built hfviewer.com to fill that gap: paste a Hugging Face model URL, open an interactive architecture graph in the browser, and move between overview and detail without installing anything.

Open Gemma 4 family page Open a sample model page Back Home

Why we are doing this

This is our way of giving back to the Hugging Face community.

Why we built it

We kept running into the same problem: a model card can tell you what a model is for, but it rarely helps you inspect the actual structure quickly. If you want to understand where the vision encoder enters, how the decoder repeats, whether the model routes through experts, or how a multimodal merge happens, you often end up reading config files, staring at code, or building your own mental graph from scattered clues.

hfviewer is meant to make that first architectural pass much faster. You can open a model directly from the Hugging Face URL, get a visual map in the browser, and then zoom from the broad system shape down into the more specific substructure that matters for understanding deployment, latency, and correctness.

What hfviewer does

Open models directly from Hugging Face

Paste a model URL or repo id and open the graph without a local setup or notebook workflow.

Switch between overview and detail

Granularity levels let you move from the high-level architecture down to more specific traced blocks and paths.

Compare model families

Family pages such as Gemma 4 let you compare multiple related models with synchronized interaction instead of isolated screenshots.

This loop shows embedl/Cosmos-Reason2-2B-W4A16-Edge2 as a natively rendered higher-resolution center-column crop at 2x speed, keeping the core granularity transition prominent while the info panel stays out of view.

A new kind of interactive blog

One of the most interesting things hfviewer enables is not just a prettier model page, but a new kind of technical article. On the Gemma 4 family page, the blog text and the graph are connected. You can read a section about a particular architectural decision, jump into the corresponding part of the graph, and then move back into the article with the surrounding context still intact.

That matters because model understanding is rarely linear. Sometimes you start from prose and need to verify it visually. Sometimes you see a node, a route, or a merge in the graph and want the editorial explanation immediately. We think that graph-to-text and text-to-graph loop is a better format for ML communication than a static diagram dropped into a long post.

Where to start

Open a familiar model such as Qwen/Qwen3.5-4B to get a feel for the main interaction model.
Jump to the Gemma 4 family page to see how the same interface can support a synchronized comparison and an editorial walkthrough.

We are releasing this because we think architecture understanding should be easier to share, easier to discuss, and easier to build on. Again: This is our way of giving back to the Hugging Face community.

Interactive model view

Model

Back Home Hugging Face

Node

Back to article

Node

Short description

Screenshots

Draw a rectangle around the incorrect part.

Email for fix notification (optional) Also send me occasional hfviewer updates, new model visualizations, and product news. You can unsubscribe at any time. Privacy Policy.

Technical overview

Understanding the Gemma 4 family

HANNES VON ESSEN APR 15, 2026

Gemma 4 is easiest to understand as one decoder-centered recipe adapted to three deployment problems. The E2B and E4B members are edge models built for tight memory and latency budgets. The 31B is the dense model for serious long-context and high-quality local or server inference. The 26B-A4B changes the economics instead, exposing far more total capacity while activating only a 3.8B subset per token. The family is therefore more useful to read by bottleneck than by parameter count alone. ¹

One decoder recipe, three bottlenecks

What keeps the family coherent is the shared attention backbone. Across the lineup, local sliding-window attention is interleaved with full global attention, and the final layer is always global. The edge models use 512-token sliding windows and 128K context, while 31B and 26B-A4B move to 1024-token windows and 256K context. That matters because it makes long context an architectural choice rather than just a larger tokenizer limit.

The expensive part of the stack is also where the main optimizations are concentrated. The global layers use unified Keys and Values and apply proportional RoPE, while the cheaper local layers keep the familiar standard RoPE regime. The point is not that every layer sees the whole sequence all the time; it is that the model restores global communication often enough to make the large window operationally meaningful. ¹

The edge models do more than add modalities

The smallest dense models are the most architecturally distinctive. E2B is listed as 2.3B effective parameters but 5.1B with embeddings, and E4B as 4.5B effective but 8B with embeddings. The difference comes from Per-Layer Embeddings, which give each decoder layer its own small embedding for every token instead of forcing a compact model to preserve all linguistic detail through one bottom-layer embedding alone. In the visible graph that extra path shows up as layer-specific text embeddings feeding a Per-layer projection that keeps token-specific text structure available deeper in the stack. ²

That design choice matters because these are also the most ambitious multimodal members at their size. All models accept image input, but only E2B and E4B add native audio, pairing roughly 150M-parameter vision encoders with roughly 300M-parameter audio encoders. In a compact multimodal decoder, reserved image and audio positions can easily erode language precision. Giving later layers a direct token-specific text signal is a clean way to preserve more of that structure. ¹

The vision path is also more flexible than a fixed square-image pipeline. The VisionEncoder preserves natural aspect ratio, uses a 2D positional scheme so height and width are represented separately, and exposes soft visual-token budgets of 70, 140, 280, 560, and 1120 tokens. At masked_scatter, projected image or audio features overwrite reserved placeholder positions in the language-side sequence. That turns visual detail into an explicit latency-quality knob: lower budgets make sense for captioning or video frames, while higher budgets are better suited to OCR, document parsing, and small text. After that replacement everything still goes through the same Decoder cycle. ¹

31B is the dense long-context member

The 31B is the cleanest dense expression of the recipe. It has 30.7B parameters, 60 layers, 1024-token sliding windows, 256K context, and a much larger ~550M vision encoder. There is no routing trick and no per-layer embedding trick here; the point is always-on capacity for long documents, repositories, codebases, and large multimodal contexts where dense quality matters more than the cheapest possible token. ¹

5 sliding + 1 full 1024-token local window 256K = 262,144 tokens

The upper grid magnifies the causal look-back band so the layer schedule stays legible. The ratio strip keeps the real scale visible: most 31B layers only read a 1,024-token local history, and every sixth layer is the expensive causal full pass that reconnects the entire 256K (262,144-token) context.

The deployment implication is straightforward. Loading the weights alone is about 58.3 GB in BF16 or 17.4 GB in Q4_0, before runtime overhead and KV cache. So 31B can be made local in quantized form, but its natural home is still a serious workstation or server GPU when long-context and dense-model headroom are the priority. ²

Position handling follows the same logic. Rotary embedding remains the positional backbone. RoPE is the mechanism that injects position into attention by rotating the query and key vectors with a position-dependent phase, so token order is represented inside the attention computation itself. Gemma 4 does not use one RoPE regime everywhere, however. Sliding-attention layers are annotated with standard RoPE, while full-attention layers are annotated with proportional RoPE. Gemma 4's proportional variant changes the rotation schedule for the long-range layers by using a much larger base period and rotating only part of the attention head dimension. The long-range layers therefore age more gracefully as sequence length grows, so the periodic full-attention passes remain useful even when the sequence is very long. The point is clear: long context is treated as an inference-systems problem as much as a modeling problem.

26B-A4B changes the cost model

The 26B-A4B asks a different question. It has 25.2B total parameters, 3.8B active parameters, 30 layers, 1024-token sliding windows, 256K context, and an expert layout of 8 active experts, 128 total experts, and 1 shared expert. Instead of sending every token through the same feed-forward path, it uses a Router to decide which Experts handle each token. The model therefore exposes more conditional capacity only where the token needs it. ¹

That makes 31B complementary rather than redundant with it. The dense model buys always-on depth. The MoE model buys conditional feed-forward capacity. The savings, however, are in active compute rather than residency: all 25.2B parameters still need to be loaded for routing, which is why the Q4_0 load footprint is still about 15.6 GB. That is close enough to 31B's 17.4 GB that 26B-A4B makes the most sense on a gaming GPU or workstation, where you want much more headroom than E4B without paying for a dense 31B-style forward pass on every token. ²

Speed, accuracy, and the new local frontier

The edge speed story is the clearest. Published device measurements put E2B at 52 GPU decode tokens per second on a Galaxy S26 Ultra, 57 on an iPhone 17 Pro, and 160 on a MacBook Pro M4 GPU. E4B lands at 22, 25, and 101 tokens per second on those same device classes. That is fast enough to make the E line feel genuinely interactive on phones and laptops rather than merely able to run locally. ³

Those speeds do come with a ceiling. On MMLU Pro and GPQA Diamond, E2B scores 60.0 and 43.4, E4B scores 69.4 and 58.6, 26B-A4B scores 82.6 and 82.3, and 31B scores 85.2 and 84.3. But the smaller models are still more capable than their size class would suggest: E4B already edges the earlier 27B dense baseline on MMLU Pro and outperforms it decisively on GPQA Diamond. That is a strong sign that the compact end of the family is not just cheaper, but genuinely better positioned on the quality curve than the previous generation. ¹

The more important comparison is the same-latency one. In a recent controlled benchmark, E4B reached 0.675 weighted accuracy at 5.458 seconds mean latency, while Qwen3-8B reached 0.322 at 5.041 seconds. E2B reached 0.493 at 4.913 seconds, versus Phi-4-reasoning at 0.427 and 4.857 seconds. That is the real win for the small models: they still give up headroom to the larger members, but they appear to deliver more accuracy at roughly the same latency as nearby alternatives. ⁴

A compact deployment snapshot based on those published figures looks like this.

Model	Best-fit hardware	Published speed / cost signal	Quality signal	What you are buying
E2B	Phone, browser, Raspberry Pi, lightweight laptop	52–57 GPU decode tok/s on flagship phones; 160 tok/s on MacBook Pro M4 GPU	MMLU Pro 60.0; GPQA 43.4; weighted accuracy 0.493 at 4.913s	Fastest local multimodal member, with the lowest reasoning ceiling.
E4B	Phone, laptop edge, offline assistant	22–25 GPU decode tok/s on flagship phones; 101 tok/s on MacBook Pro M4 GPU	MMLU Pro 69.4; GPQA 58.6; weighted accuracy 0.675 at 5.458s	The strongest small-model operating point.
26B-A4B	Gaming GPU, RTX-class workstation, high-end local box	3.8B active params per token; ~15.6 GB Q4_0 load	MMLU Pro 82.6; GPQA 82.3	Much more conditional capacity without dense 31B compute every pass.
31B	Workstation or server GPU	30.7B dense model; ~17.4 GB Q4_0 load or ~58.3 GB BF16 load	MMLU Pro 85.2; GPQA 84.3	Maximum dense quality and long-context headroom.

Values are a deployment snapshot rather than a single apples-to-apples benchmark across one toolchain and one hardware stack.

What Gemma 4 means

What makes the family interesting is that its members are not just scaled copies of one another. The E2B and E4B edge models add per-layer text scaffolding and audio because compact multimodal decoders need extra help preserving language precision under edge constraints. 31B stays dense because long-context quality benefits from always-on capacity. 26B-A4B uses routing because a workstation can often afford the residency of a larger model even when it cannot afford to spend dense 30B-class compute on every token. ⁵

That gives practitioners a cleaner selection rule than parameter count alone. Choose E2B or E4B when privacy, battery, latency, and offline multimodality are the constraint. Choose 26B-A4B when you have a gaming GPU or workstation and want a better capacity-per-token bargain. Choose 31B when dense quality and long-context reliability are worth building heavier hardware around. The family's real contribution is that each member moves a different bottleneck while still sharing the same architectural center. ¹

Architecture graph for embedl/Qwen3.5-0.8B-FlashHead.

Opening your dashboard.

Visualize your model repos.

Model dashboard.

Not ready

Generating

Ready

Featured

Featureyour model

Manual README.md edit

README.md preview

README.md

Seventy quote block variants.

Prism rail

Folded glass

Signal frame

Editorial shell

Graph glow

Layer stack

Insight capsule

Aurora card

Blueprint

Minimal pro

Margin rail

Corner field

Underline pulse

Diagonal shard

Terminal line

Node path

Manuscript mark

Theorem strip

Waveform

Glass cutout

Dot anchor

Orbit dot

Node chain

Pinned insight

Constellation

Beacon

Split rail dot

Endpoint dots

Spotlight dot

Inline dot badge

Dot anchor clean

Dot halo

Left glow wash

Dot notch

Inline leader dot

Floating pin

Soft dot field

Blue dot anchor

Dot with echo

Thesis mark

Thesis soft glow

Wide soft field

Small breathing mark

Left bloom

Quiet breath

Blue-green field

Edge thesis dot

Double breath field

Top note

Underline glow thesis

Unclipped thesis soft glow

Unclipped wide soft field

Unclipped small breathing mark

Unclipped left bloom

Unclipped quiet glow

Unclipped blue-green glow

Unclipped edge spark

Unclipped double breath

Unclipped top note

Unclipped underline glow

Static thesis glow, card-clipped

Static wide field, card-clipped

Static small mark, card-clipped

Static left bloom, card-clipped

Static quiet glow, card-clipped

Static blue-green glow, card-clipped

Static edge spark, card-clipped