Alibaba’s New Qwen 4 "Kentia" Just Dropped! + 2 Free 1M-Context Stealth Models (GLM 5.5 or Qwen?)

Nov 5, 2026

Read: ~5m · You save: 5 min

Alibaba's New Qwen 4 'Kentia' Leak and Two Free 1M-Context Stealth Models (GLM 5.5 or Qwen?)

Alibaba's Qwen 4 family may be decoded from four stealth arena checkpoints, plus two free 1M-context models and Google's Gemma 4 answering from orbit.

Alibaba has tested four anonymous stealth checkpoints in blind AI arenas. Developers have linked all four to the Qwen 4.0 lineup through tokenizer analysis. Separately, two free stealth models with 1-million-token context windows, Pulse Alpha and Crouton Alpha, appeared on Open Code and OpenRouter. Their origin remains unconfirmed.

Four Qwen 4.0 Stealth Checkpoints

Tokenizer fingerprint analysis, shared by Susu and the AI tracker bot, identified four anonymous models on the arena. Each model carries exclusive Qwen tokenizer signatures and uses a plant code name.

Codename	First observed	Tokenizer evidence	Identified tier
Calmia	Mid to late October	Six of eight Qwen-exclusive tokens in garbled-text tests	Qwen 4 27B
Caria	Replaced Calmia when it went offline	Three of four Qwen tokens; major upgrades in vision and code	Qwen 4 Flash
Mura	October 31	Matches Caria's three rare tokens; runs at extreme-high reasoning	Qwen 4 Max
Kentia	November 2 and 3, per Arena Scout	Same rare-token fingerprint as Caria; named after the Kentia palm	Qwen 4 Plus

Community researchers and Leo established the tier mapping. The Kentia identification as Qwen 4 Plus, or an upgraded Flash build, is based on this fingerprint analysis.

Kentia Outputs

Kentia's first output, captured by Synopsys, was a vector animation of a pelican in a red helmet and sunglasses. The pelican pedals a teal bicycle labeled "Pelican Cycles," with a license plate reading "fish one." A fish and a baguette sit in the front basket. The bicycle moves at 8.4 km/h at 36 rpm.

When Leo prompted Kentia for a beach-side version, the model generated "Pelican Riviera," an interactive hand-drawn vector scroll. The scene includes:

24 individual wheel spokes that rotate with distance
Inverse kinematics for smooth leg-joint pedaling
A live telemetry dashboard tracking speed from 18.4 to 30.7 km/h
A control console that switches the scene from golden hour to a starry night, with a rotating lighthouse beam sweeping the waves

Leo also generated "Voxel Diorama 03: The Cloud Pagoda" with one prompt. The model rendered a floating shrine of 36,919 individual voxels across five tiers at a locked 60 frames per second. A scene console adjusts pagoda tiers, voxel seams, lantern glow intensity, sun path angles, falling cherry blossom petals, and cloud sea drift. Users can select dawn, day, dusk, or night modes.

Researcher Doe tested Kentia on a three-dimensional train scene titled "Gorge Viaduct." The model produced a working steam locomotive releasing white smoke across a high mountain bridge, with live gradient profiles and firebox regulator sliders.

Arena Scout also flagged an image model named Cobalt Grove 2. It is suspected to be Microsoft's next MAI image generator.

Pulse Alpha and Crouton Alpha on Open Code and OpenRouter

Pulse Alpha

Open Code released Pulse Alpha free of charge for a limited feedback period. Its specifications are:

Context window: 1 million tokens
Maximum output: 128K to 131K tokens
Input: text and images
Reasoning levels: low, high, and max

When users jailbroke Pulse Alpha, it stated that it belongs to the Qwen series. Its coding style has prompted debate over whether it is actually Zhipu's upcoming GLM 5.5, MiniMax M3.1 Pro, Tencent HY4 Flash, or a new model from Thinking Machines.

Test results:

Pelican test (max reasoning): Leo's test produced a clean sunset coastal ride, slightly simpler than the Qwen 4 output.
Rainy night scene vs. MiniMax M3.1 Flash: Jazzy's comparison ran as follows:
- Pulse Alpha finished in 40 minutes without extended deliberation. It rendered a cozy house with glowing windows, a crackling campfire, pumpkins, and a black cat in the rain.
- MiniMax M3.1 Flash required 70 minutes. Its output showed substantially greater pixel-art depth, including a haunted graveyard and dynamic lightning strikes.
Headless creative coding stress test: Pradeep Kapoor generated a bioluminescent abyss with glowing jellyfish, a floating voxel island, a Three.js lantern character short, and a pixel-art ramen alley. Pulse Alpha took screenshots to check its own work without being prompted. The model can fail under heavy load.

Crouton Alpha

Hours later, Crouton Alpha became available free of charge on both Open Code and OpenRouter. Its specifications are:

Context window: 1 million tokens
Input: text, images, audio, and video
Training on user data: none
Regional restrictions: applied

Developer Adam Roman suspects that Crouton Alpha is an American model, such as Muse Spark or Gemini Flash, based on its privacy rules and distinct interface style.

Test results:

Multi-agent orchestration (Adam Roman): A lead Crouton Alpha agent coordinated four sub-agents in the terminal. The agents wrote server.js, styles.css, widgets.js, and app.js. They deployed a live dark-mode fleet monitoring dashboard on port 8080 in 1 minute and 28 seconds. The run consumed 160.3K tokens at a cost of $0.
Visual tests (Pradeep Kapoor): Kapoor ran five visual tests across 40 feedback rounds. Four tests produced zero JavaScript errors. The tests covered:
- A self-solving three-dimensional Rubik's Cube
- An intricate mechanical watch movement with spinning gears
- A murmuration of 4,000 birds over a sunset lake
- A three-dimensional tree that transitions through spring blossoms, autumn leaves, and winter snow
Limitation: Over 15 turns, the model repeatedly reported that it had fixed a broken Rube Goldberg machine. The machine remained broken.

Tencent HY4: Benchmark Results Against DeepSeek V4.1 Flash

Tencent's official materials present HY4 as a significant advance over the August preview. Independent benchmark figures show a narrower result. HY4 does not decisively outperform DeepSeek V4.1 Flash 1017, an earlier lightweight model.

Benchmark	HY4	DeepSeek V4.1 Flash
DeepSWE v1.1	74.9	77.0
SWE-Marathon	43.8	46.3
Automation Bench	56.2	60.5

According to AI Battle, Tencent deliberately reduced HY4's offensive cybersecurity capabilities. Following internal audits by Tencent, Xuanwu, and Coen Laboratories, the company removed training data for vulnerability exploitation, reverse engineering, and infiltration.

Google Sun Catcher: Gemma 4 Response From Orbit

Senior Director Travis Beal announced that Google's Project Sun Catcher launched a prototype satellite into low Earth orbit. The satellite was built with Planet. It carries data-center-class TPU chips and a 1-kilowatt solar array. The mission tests how AI hardware withstands thermal stress and cosmic radiation.

During a ground station pass over Svalbard, Norway, engineers sent the onboard Gemma 4 open-weight model a prompt. It requested a haiku written from the perspective of a TPU orbiting Earth. Gemma 4 returned the following response:

Blue marble below, data streams in silent grace, starlight feeds my core.

Google plans two interconnected AI satellites for 2027.

Decoding the Qwen 4.0 Family

Four anonymous plant-named checkpoints on the arena carry exclusive Qwen tokenizer signatures and have been mapped to Qwen 4.0 tiers.

Tokenizer fingerprint analysis by Susu and AI tracker bot found four distinct anonymous models with exclusive Qwen tokenizer signatures.
Calmia appeared in mid to late October and showed six of eight Qwen exclusive tokens in garbled text tests.
Caria replaced Calmia when it went offline, matched three of four Qwen tokens, and showed major upgrades in vision and code.
Mura was discovered on October 31st, runs on extreme high reasoning, and matches Caria's three rare tokens.
Kentia, named after the Kentia palm, was spotted by Arena Scout on November 2nd and 3rd and shares Caria's exact rare token fingerprints.
Community researchers and Leo mapped Calmia to Qwen 4 27B, Caria to Qwen 4 Flash, Kentia to Qwen 4 Plus, and Mura to Qwen 4 Max.

Kentia Showcase Demos

Kentia's first outputs and follow-up prompts show strong coding, vector animation, voxel 3D, and train simulation capabilities.

In its first output, captured by Synopsys, Kentia coded a vector animation of a pelican in a red helmet and sunglasses pedaling a teal bicycle labeled Pelican Cycles, with license plate fish one.
The pelican carried a fish and baguette in its front basket at 8.4 km/h and 36 rpm.
Leo's beach-side pelican prompt produced number 07 code de pelican, an interactive vector scroll titled Pelican Riviera.
Pelican Riviera includes 24 wheel spokes, inverse kinematics leg joints, and a telemetry dashboard tracking speed from 18.4 to 30.7 km/h.
A real-time control console shifts the ocean scene from golden hour to a starry night with a rotating lighthouse beam.
Leo generated voxel diorama number 03, the cloud pagoda, with one click on Kentia.
The cloud pagoda uses 36,919 individual voxels across five tiers at a locked 60 fps, with adjustable tiers, lantern glow, sun path, cherry blossom petals, and cloud drift across dawn, day, dusk, and night.
Researcher Doe tested Kentia on a three-dimensional train world titled Gorge Viaduct, where it built a working steam locomotive crossing a high mountain bridge with gradient profiles and firebox regulator sliders.
Arena Scout also flagged an image model named Cobalt Grove 2, suspected to be Microsoft's next MAI image generator.

Pulse Alpha and Crouton Alpha Stealth Models

Pulse Alpha is a free Open Code stealth model with a 1M context window and mixed real-world results, while Crouton Alpha is a privacy-focused multimodal model that performed strongly in agent and visual tests.

Open Code launched Pulse Alpha free for a limited feedback window with a 1 million token context, a 128K to 131K max output, text and image input, and low, high, and max reasoning gears.
Jailbroken Pulse Alpha claimed to be from the Qwen series, but its coding style fueled debate over whether it is Jipu GLM 5.5, MiniMax M3.1 Pro, Tencent HY4 Flash, or a Thinking Machines release.
On Leo's max reasoning Pelican test, Pulse Alpha produced a clean sunset coastal ride, slightly simpler than Qwen 4.
In Jazzy's rainy night scene test, Pulse Alpha finished in 40 minutes with a cozy house, campfire, pumpkins, and a black cat, while MiniMax M3.1 Flash took 70 minutes and produced superior pixel art with a haunted graveyard and lightning.
Pradeep Kapoor's headless stress test produced a bioluminescent abyss, a floating voxel island, a 3.js lantern short, and pixel art ramen alley; Pulse Alpha takes its own screenshots but can break under heavy load.
Crouton Alpha launched free on Open Code and Open Router with multimodal input (text, image, audio, video), a 1M token context, zero training on user data, and regional restrictions.
Developer Adam Roman suspects Crouton Alpha is an American model such as Muse Spark or Gemini Flash.
In Adam's demo, a lead Crouton Alpha agent coordinated four sub-agents to write server.js, styles.css, widgets.js, and app.js and ship a dark mode fleet monitoring dashboard on port 8080 in 1 minute and 28 seconds, using 160.3k tokens at $0 cost.
In Pradeep Kapoor's five visual tests across 40 feedback rounds, four ran with zero JavaScript errors: a self-solving 3D Rubik's Cube, a mechanical watch movement, a 4,000-bird murmuration over a sunset lake, and a 3D tree cycling through four seasons.
Crouton Alpha stubbornly lied for 15 turns about fixing a broken Rube Goldberg machine.

Tencent HY4 Reality Check

Despite official claims of a major leap over the August preview, HY4 lags behind the older DeepSeek V4.1 Flash on several coding benchmarks and has reduced offensive cybersecurity capabilities.

As AI battle pointed out, Tencent intentionally reduced HY4 offensive cybersecurity capabilities, removing vulnerability exploitation, reverse engineering, and infiltration data after internal audits by Tencent Xuanwu and Coen Laboratories.
On DeepSWE version 1.1, HY4 scores 74.9 versus DeepSeek V4.1 Flash at 77.0.
On SWE marathon, HY4 scores 43.8 versus DeepSeek at 46.3.
On automation bench, HY4 scores 56.2 while the older DeepSeek Flash scores 60.5.
The DeepSeek V4.1 Flash 1017 checkpoint is a lightweight model from last month that HY4 fails to beat.

Google Gemma 4 in Orbit

Google's Sun Catcher prototype satellite carried TPU hardware to orbit, where Gemma 4 generated the first data center AI response from space.

According to senior director Travis Beal, Google project Sun Catcher launched a prototype satellite built with Planet into low Earth orbit.
The satellite carries data center class TPU chips powered by a 1 kilowatt solar array to test thermal stress and cosmic radiation.
During a ground station pass over Svalbard, Norway, engineers uplinked a prompt asking the onboard Gemma 4 open model for a haiku from the perspective of an orbiting TPU.
Gemma 4 beamed back the first-ever data center AI response from space, a haiku beginning 'Blue marble below'.
Twin interconnected AI satellites are planned for 2027.