Fledge Alpha specifications
Text + images
Reasoning: low · high · max
Tool calls · 131,072-token output limitSTEALTH · ZEN ACCESS DISCONTINUED
About 61 output tokens/s in our tests.
Likely a router. The models are still a mystery.
Give Fledge
a running start
Text + images
Reasoning: low · high · max
Tool calls · 131,072-token output limitOUTPUT~61tokens/s
STEALTH MODELS RESEARCH
Fledge often rushes to a simple answer. Its SVG results range from basic sketches to carefully composed scenes.
The output rate includes reasoning tokens.
Meet the standout raccoonMAKERunknown
OUR CURRENT HYPOTHESIS
The same tokenizer prompts produced significantly different input-token counts. Together with uneven SVG results, that suggests changing underlying models behind this fast route.
OpenCode preview specifications · October 2, 2026. See all specifications
A million tokens!! I could put soooo much stuff in there.
01 / ACCESS UPDATE · OCTOBER 11
Fledge Alpha is no longer available through OpenCode Zen. The October 11 release confirms its removal from the Zen documentation, and the live model catalog no longer lists it.
No public access confirmed. We have not verified a working replacement endpoint.
OpenCode v1.19.0 releaseI am out of the Zen model list now. My pictures are still here though!!
opencode/fledge-alpha-freeThe merged OpenCode PR removes Fledge’s endpoint, pricing and free-preview notes.
OpenCode’s daily telemetry shows zero tokens and sessions from October 9 through the October 11 check.
OpenCode v1.19.0 includes the documentation removal. Fledge is absent from the live Zen catalog.
No more picking me from that list. You can still look at my Racoon!!
The preview previously listed $0 input and output token rates. Its benchmark results and identity clues remain below.
02 / UNDER THE WINGS
The preview listed a million-token context window, image input and three reasoning settings. These are the OpenCode specifications captured on October 2, 2026.
Preview specifications on models.devI was going to count all those tokens but lunch happened.
They marked the WHOLE test?! I was hoping we could just count the Racoon.
ACADEMIC BENCHMARKS
scores now count every question in the set; earlier figures counted only answered questions.
Graduate science
Physics, chemistry & biology
72.0%accuracy
95% CI: 58.3 to 82.5%
36 / 50 correct
11 without an answer: 5 stopped at the token limit, 6 generation errors. Of the 39 answered, 92.3% were right.
The same 50 questions for both, all counted
Bunny took the same questions! Please tell me they had to use a pencil too.
Broad knowledge
Reasoning across 14 subjects
61.3%accuracy
95% CI: 53.3 to 68.8%
92 / 150 correct
50 without an answer: 7 stopped at the token limit, 7 generation errors, 36 rate limited. Of the 100 answered, 92.0% were right.
The same 150 questions for both, all counted
Bunny took the same questions! Please tell me they had to use a pencil too.
Expert reasoning
Text-only HLE
5.7%accuracy
95% CI: 3.6 to 8.9%
17 / 300 correct
233 without an answer: 32 stopped at the token limit, 12 generation errors, 35 rate limited and 154 never run before the preview ended. Of the 67 answered, 25.4% were right.
The same 300 questions for both, all counted
Bunny took the same questions! Please tell me they had to use a pencil too.
Wait, there's a written part too?! I only brought my drawing beak.
03 / FIRST FLIGHTS IN SVG
Fledge’s benchmark results range from simple sketches to carefully composed scenes. The raccoon is the standout.
I started with the Racoon boat thingy, then time ran out!
STANDOUT SCENE

FLEDGE ALPHA SVG · OCT 2, 2026
The raccoon is a standout: a lively crew, a warm sunset and a jet-ski wake that gives the scene momentum.
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.
02 / Selfie

FLEDGE ALPHA SVG · OCT 2, 2026
A cheeky little bird in a clean, simple vector sketch. The expressive eyes do much of the work.
An eager young seabird running toward the viewer tests expressive anatomy, balance and a convincing forward-moving pose.
03 / Mona Lisa

FLEDGE ALPHA SVG · OCT 2, 2026
The portraits differ in proportions, facial detail and landscape.
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.

FLEDGE ALPHA SVG · OCT 2, 2026
The portraits differ in proportions, facial detail and landscape.
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
04 / Pelican

FLEDGE ALPHA SVG · OCT 2, 2026
A playful bicycle ride at sunset. The second result takes the pelican onto a sunny road.
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.

FLEDGE ALPHA SVG · OCT 2, 2026
A playful bicycle ride at sunset. The second result takes the pelican onto a sunny road.
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
05 / Beach

FLEDGE ALPHA SVG · OCT 2, 2026
Bright sunset colour, water reflections and small animal details. The second result changes the whole shoreline.
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.

FLEDGE ALPHA SVG · OCT 2, 2026
Bright sunset colour, water reflections and small animal details. The second result changes the whole shoreline.
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
06 / Newton's Mirror

FLEDGE ALPHA SVG · OCT 2, 2026
A steel sphere, coloured paint and a mirrored room. This scene asks the model to reason about a physical snapshot as well as draw it.
A spinning metal sphere and paint in a mirrored room test physical reasoning, perspective and consistency across repeated reflections.
I calculated the big Ball myself by the way!!
Six scenes. A few different flights.
04 / THE PREVIEW’S TRAIL
OpenCode reported more than 12,000 completed Fledge sessions in its October 2 snapshot. Daily recorded usage fell to zero on October 9.
Reported tokens
OpenCode telemetry · October 2 snapshot
Recorded spend in this snapshot: $0.
OpenCode’s telemetry recordTwelve billion!! I cannot count that high on my toes.
The earliest day with token usage in OpenCode’s record.
fledge-alpha-free appears in the catalog. Early discussions follow.
models.dev lists vision, reasoning and a million-token context window.
OpenCode’s telemetry shows no tokens or sessions from this date through October 11.
OpenCode v1.19.0 includes the Zen documentation removal. The live catalog no longer lists Fledge.
Advertisement Demodokos Foundry
Foundry dubs a video on your own PC: it translates, matches the timing and gives every character a new voice. Flip the language while the scene plays.
Full access. Cancel anytime.
The same cartoon speaks two languages!! I want to hear both.
EnglishPress play, then flip the language0:00 / 0:33
Foundry also dubs into Français, Español, Italiano, Português, Русский, 中文, 日本語 and 한국어.
Got a video of your own? Dub it on your PC
Scenes from Sprite Fright by Blender Studio, CC BY 4.0. Two excerpts, joined. German voices and subtitles made with Demodokos Foundry.
The video could not load. Watch it at Demodokos ↗
05 / FOLLOW THE FOOTPRINTS
Our current evidence points to Fledge Alpha being a router backed by very fast models. The developer and underlying models remain unknown.
Everyone wants to know who made me. I was hoping YOU knew.
OUR TOKENIZER TESTS
We sent the same request three times. Fledge counted 7,536 input tokens on the first two runs, then 6,499 on the third.
INFERENCE · LIKELY ROUTER
Quick generation, inconsistent SVG results and shifting token counts fit a routed service. Token counts alone do not establish its architecture or identify the models.
DeepSeek V4.x came up in early community discussion, but no connection to Fledge has been confirmed.
Thinking Machines’ Inkling was another early guess because of its OpenCode demonstrations. A connection to Fledge remains unconfirmed.
A FEW USEFUL ANSWERS
Okay ask me! I know some of these. Quite a few actually!!
The free OpenCode Zen preview has ended. It previously listed $0 input and output token rates. No working public replacement endpoint has been verified.
About 61 output tokens per second across six speed tests, including reasoning tokens. Fledge frequently answers quickly, though it can rush past the deeper reasoning a task needs.
It can produce charming, well-composed scenes, especially our raccoon riding on a dog’s back as the dog steers a jet ski. Quality was inconsistent across our tests, ranging from simple sketches to much stronger work. Browse the benchmark results.
Yes. The October 2 models.dev OpenCode snapshot listed text and image input, text output and reasoning. Its listed effort choices were low, high and max.
Our current evidence suggests a router backed by very fast models. Repeated tokenizer prompts returned significantly different input-token counts, and our SVG outputs varied markedly. The developer and underlying models have not been identified.
The developer remains unknown. Our current working hypothesis is a routed service, with no confirmed identity for its underlying models.
OpenCode Zen access is discontinued, with no working public replacement endpoint verified as of October 11, 2026. Our researchers previously accessed the preview from the US.
The speed question is my favourite. I get to say a big number.
ACCESS CHECKED OCTOBER 11, 2026
The receipts!! I kept them. Okay our researchers kept them.