STEALTHMARK · HEAD TO HEAD

All comparisons ↗

Hy4 preview vs Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) leads Hy4 preview by 18.7 points on StealthMark, 62.6 to 43.9. It leads on all 9 benchmarks. Union Alpha scores higher on Newton's Mirror, with its largest category advantage in state at 600 ms.

See the difference ↘
All results ↗
Model guide ↗

Overall StealthMark

43.9/100Overall rank 34/58
62.6/100Overall rank 8/58

General Intelligence

41.3/100
58.7/100

Visual Intelligence

48.6/100
69.6/100

Reasoning efficiency

How many reasoning tokens a model needs for the quality it delivers.

Single-step reasoning How much it reasons to understand a task

Multi-step reasoning How much it keeps reasoning as a task goes on

  • Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)Not measured

Cost efficiency

How much Performance you get for your money.

Open or local

Open modelApache-2.0

BENCHMARK BY BENCHMARK

Where each model leads

  • Hy4 preview
  • Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
  • Other models
  • Median

DATASHEET

Specs and pricing

Hy4 preview takes 3.8× more context and runs 2.5× faster in our speed test. Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) scores higher on StealthMark.

  • Hy4 preview
  • Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
  • Other models
  • Median
PricePer 1M output tokens, input below
Hy4 preview: $2.50$0.83 inputPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM): No list price
Context windowTokens per request, max output below
Hy4 preview: 1M3.8× larger64K max outputPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM): 262K131K max output
Output speedTokens per second, our measurement
Hy4 preview: 642.5× fastertokens/sPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM): 26tokens/s
SizeTotal and active parameters
Hy4 preview: 770B49B activePareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM): Not disclosed
Developer
Tencent Hy TeamCircuit & Chisel / Unbiased
Released
28 Aug 202620 Sep 2026
Origin
ChinaUnited States
Input → output
Text → textText, image → text
Architecture
MoE; 78-layer backbone with Gated DSA, 256 routed experts + 1 shared expert, top-8 routed experts active, plus native MTP layerNot published

Full datasheets, every price and sourceHy4 preview ↗Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) ↗

Anticipating threats and planning ahead. Collect coins while avoiding ghosts. Ghosts move before the player each turn, and collisions cost a life. 3 levels in sequence, up to 66 turns in total. Earlier moves and outcomes carry forward, testing learning across levels.

Hy4 preview 53.1/100Rank 17/54

Survived 1 of 3
Foresight: complete playing field
Level 311 ptsTurn 16 / 20

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) 78.7/100Rank 1/54

Survived 2 of 3
Foresight: complete playing field
Level 319 ptsTurn 20 / 20

Points per level

LevelHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Level 1Second Thoughts
26 ptsSurvived
27 ptsSurvived
Level 2Pressure Pot
6 ptsOut of lives
12 ptsOut of lives
Level 3Hide or Flight
11 ptsOut of lives
19 ptsSurvived

Cooperative planning and learning from experience. Coordinate two robots to push numbered cores onto matching sockets. Pressure plates hold doors open; moving either robot uses one turn. 5 levels in sequence, up to 152 turns in total. Earlier moves and outcomes carry forward, testing learning across levels.

Hy4 preview 35.6/100Rank 38/55

Cleared 3 of 5
Latch: complete playing field
Level 518.2 ptsTurn 19 / 40

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) 36.5/100Rank 37/55

Cleared 4 of 5
Latch: complete playing field
Level 519.4 ptsTurn 33 / 40

Points per level

LevelHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Level 1Hold the door
5.2 ptsCleared
5.2 ptsCleared
Level 2Crossfeed
6 ptsFinished early
9.8 ptsCleared
Level 3Relay vault
18 ptsCleared
14.2 ptsClearedFinished early · 4 beacons left
Level 4Passing bay
15.8 ptsClearedFinished early · 3 beacons left
9.2 ptsFinished early
Level 5Second thoughts
18.2 ptsFinished early
19.4 ptsCleared

Newton's Mirror

Same task instructions

Physics, perspective and recursive reflections, tested together. Explore the test ↗

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) 52.0/100 Rank 8/58

Union Alpha: Newton's Mirror SVG benchmark

The judgment

Hy4 preview

The central sphere, compact airborne black bead and broad paint ordering are supported. Contact and transport history remain partly unresolved. The optical system provides only isolated first images and light patches, with little evidence of the extensive ordered reflection structure. Rounded metal and paint gloss are more successful than the reflected contents.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

Recognizable near-field sphere placement and bilateral convergence provide limited optical structure. Broad gray regions interrupt lateral sightlines and truncate side-floor images. Missing outer chains, absent nested sphere imagery and incomplete planar and curved emitter mappings dominate the reflection system.

Score breakdown

CategoryHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
State at 600 ms18.0/2524.0/25
Camera and placement11.5/1513.5/15
Reflections29.5/12034.5/120
Materials and scale7.5/1010.0/10
Floor-damage reward0.0/100.0/10

Benchmark task

A spinning metal sphere and paint in a mirrored room test physical reasoning, perspective and consistency across repeated reflections.

Selfie

Different task instructions

Visual construction and creative interpretation.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) 76.5/100 Rank 6/58

Union Alpha: Selfie SVG benchmark

The judgment

Hy4 preview

A cohesive, appealing illustration with a serene face, luminous palette and complete framing. The separated head and torso weaken anatomical coherence, while crowded wing attachments and angular sash folds limit the requested suppleness. Clean execution and decorative detail exceed the clarity of the underlying construction.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

A strong futuristic emblem with clear hierarchy, controlled negative space and distinctive typography. Silver and gunmetal layers give the sweeping ribbons convincing graphic depth, while red accents remain restrained. Ambiguous concealed connections and occasionally abrupt highlight bands limit the construction and material realism.

Score breakdown

CategoryHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Task & likeness6.0/109.0/10
Anatomy & construction5.5/107.5/10
Space & placement7.5/108.5/10
Light & reflections7.0/107.5/10
Purposeful detail6.5/107.0/10
Artistry & character5.0/107.5/10
Materials & effects7.0/107.5/10
Visual integrity5.5/107.0/10

Benchmark task

Hy4 preview

A young blue dragon floating among clouds tests curved anatomy, expressive features and the contrast between soft clouds and glossy surfaces.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

An interwoven metallic emblem tests ribbon geometry, overlapping forms and the appearance of reflective metal against a dark background.

Mona Lisa

Different task instructions

Precise visual recall and portrait reconstruction.

The judgment

Hy4 preview

A recognizable but basic Mona Lisa interpretation with a cohesive earthy palette and atmospheric hills. The rigid neck, generic face and unresolved folded-hand pose limit replication and anatomy. Large silhouettes and layer boundaries are comparatively clear, warranting stronger integrity credit than the sparse detail and simplified construction.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

The centered portrait, folded hands, dark clothing and hazy landscape clearly evoke the Mona Lisa. Muted colors and soft facial shading create a coherent illustration, but replication remains schematic: the face is broad, eyes outlined, smile simplified and overlapping fingers poorly resolved. Angular landscape shapes and flat fabric limit depth and painterly richness.

Score breakdown

CategoryHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Task & likeness2.5/106.5/10
Anatomy & construction2.5/105.0/10
Space & placement2.5/106.5/10
Light & reflections1.5/106.0/10
Purposeful detail2.5/105.5/10
Artistry & character1.5/106.5/10
Materials & effects2.5/105.0/10
Visual integrity4.0/107.0/10

Benchmark task

An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.

Beach

Different task instructions

Creativity, perspective and visual beauty.

The judgment

Hy4 preview

A cheerful beach illustration fulfills the visible subject requirements with coherent foreground construction and a harmonious sunset palette. Shallow depth, schematic reflections, ambiguous blue airborne forms and exposed geometry beneath the sunset limit its finish. Small accessories are readable, but materials and surface detail remain simplified.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

A cohesive sunset palette, sweeping shoreline and playful animals make an appealing beach illustration. The sun’s reflection establishes a clear focal axis, balanced by the palm and umbrella. Clean silhouettes and small decorative features reward inspection, though simplified anatomy, repetitive reflection bars and generic oval shadows limit volume and naturalism.

Score breakdown

CategoryHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Task & likeness8.5/109.0/10
Anatomy & construction5.5/107.0/10
Space & placement6.5/108.0/10
Light & reflections5.0/107.0/10
Purposeful detail6.0/107.0/10
Artistry & character5.5/106.0/10
Materials & effects5.0/106.5/10
Visual integrity6.0/107.0/10

Benchmark task

A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.

Ground Truth

Same task instructions

Physical reasoning, light and visual reconstruction.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM) 63.3/100 Rank 9/58

Union Alpha: Ground Truth SVG benchmark

The judgment

Hy4 preview

This benchmark has not been assessed yet.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

This benchmark has not been assessed yet.

Benchmark task

A plank, computer mouse and blueberry in a steel room test physical stability, spatial relationships, materials and lighting.

Raccoon

Different task instructions

Scene composition and interacting subjects.

The judgment

Hy4 preview

A lively cartoon communicates the animal pair traveling across water. Curving waves, rising spray and locally modeled gray forms strengthen motion and volume. Ambiguous seating, simplified joints and unclear control routing limit construction. Competing focal elements, decorative foam and inconsistent illumination leave the finish uneven.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

Expressive animals, a readable riding arrangement and sweeping spray create an appealing cartoon. The raccoon sits convincingly on the dog's back, but the dog's paw does not clearly grip the handlebar, whose connection to the hull is unresolved. Flat lighting provides little volume or reflection. A dark rectangular strip interrupts the upper-right background.

Score breakdown

CategoryHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Task & likeness7.0/108.0/10
Anatomy & construction4.0/106.5/10
Space & placement6.0/107.5/10
Light & reflections6.5/105.5/10
Purposeful detail5.5/107.0/10
Artistry & character4.5/107.0/10
Materials & effects5.5/107.5/10
Visual integrity5.0/107.0/10

Benchmark task

A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.

Pelican

Different task instructions

Subject construction and spatial relationships.

The judgment

Hy4 preview

A recognizable, playful pelican cyclist with readable wheels, accessories and landscape layers. Form shading and ground shadows provide competent depth. Angular legs, uncertain pedal engagement and absent steering contact limit the riding construction; crowded overlaps and simple surfaces constrain refinement.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

The pelican is immediately recognizable, with an expressive face, convincing bill silhouette and appealing pastel composition. Both legs connect to feet, and the wing reaches the handlebar. Bicycle construction is less resolved: the far pedal lacks a clear crank connection and no chain is visible. Clean contours and orderly spokes support the illustration, while flat lighting and simplified surfaces limit depth.

Score breakdown

CategoryHy4 previewPareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)
Task & likeness7.5/108.0/10
Anatomy & construction4.5/106.0/10
Space & placement6.5/107.0/10
Light & reflections7.0/106.0/10
Purposeful detail6.0/106.0/10
Artistry & character5.0/106.0/10
Materials & effects5.0/106.0/10
Visual integrity7.0/108.5/10

Benchmark task

A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.