quietmoney

Modelflare

Low-cost LLM API relay: one endpoint to call GPT, Claude, Grok, Gemini and image/video models, with per-request usage and cost logging.

modelflare.dev ↗AI platform / APIChina 43.91%, United States 34.69%, Hong Kong 9.27%#347 of 606
An LLM API relay run by a US + Singapore team under the US shell company Havenbyte LLC; domain registered 2026-06, launched three months ago It claims to have "operated for over a year, originally serving overseas enterprise clients". After pivoting to individual developers two months ago, team members posted directly in V2EX's promotion board (disclosing their "conflict of interest"), using "$0.25 on sign-up, another $0.5 for commenting, 0.1x GPT rate, linux.do login" to break into the Chinese freebie-hunting developer scene, then turned users into promoters with "8% of first top-up credit for both inviter and invitee + long-term cash commissions for promoters + a project sponsorship plaza". Traffic is 67% direct, 25% referral, with almost no search.One-line growth story, compiled by us from public sources; speculation is marked separately below.
Open Modelflare ↗

Traffic model, last monthFrom first hearing about it to reaching the checkout page. Circles show how many passed through last month

Grey balls: target audience, not users yetGreen balls: Users (returning users who come back to top up and check usage, or people who kept browsing after the visit)One ball ≈ 5,807 people (100 balls every 5 seconds stand for one month)Click a circle for details and source links
Where the numbers in this chart come from

The main engine is referral commissions: both sides of an invite get 8% credit on the first top-up, and promotion partners earn cash commissions; direct visits are 67%, referrals are second, and search is minimal.

The turbo boost is promotion posts on V2EX: team members posted twice in a row in the promotion section, giving sign-up credit, posting redemption codes and offering GPT at a 0.1x rate, breaking into the Chinese bargain-hunting developer community.

By inference, about half of visits are existing users coming back to top up and check usage; visit counts are estimates, and about 7 of every 100 visits click through to the checkout page.

Model: Lenny's racecar growth framework. Engine = referral commissions (both sides of an invite get 8% credit on the first top-up, cash commissions for promotion partners, links placed through project sponsorship); turbo boost = the credit posts in the V2EX promotion section, breaking into the Chinese bargain-hunting developer community; lubricant = low-rate groups and top-up thresholds, with existing users coming back to top up for the cheap prices.

Real data: Biggest channel direct at 67.01%, top three channel ranking (direct, referrals, organic search), organic search 86.69% and paid search 13.31% of search, 78 search keywords in total, top search keywords (modelflare, grok api, deepseek4.0 version-exp, qwen 3.8 mgsm, gpt6 vs fable), top referrals (veridrop.org 3.07%, z2u.com 0.78%, v2ex.com 0.7%), referring site category (programming and developer software 35.95%), countries (China 43.91%, US 34.69%, Hong Kong 9.27%, Singapore 6.26%, Japan 3.75%), bounce rate 24.96%, 7.05 pages per visit, average time on site 8 minutes 31 seconds, global rank #290895, all from the Similarweb public page for last month; the public page does not give total visits or the shares of the second and third channels. 8,596 visits (about 8,600) to a Stripe checkout page are from quietmoney payment page data.

Supporting evidence: The site footer shows Havenbyte LLC in the US, and the V2EX poster says the team is in the US and Singapore; the domain was registered in 2026, and X @modelflare_dev started posting around the same time; team members posted twice in a row in the V2EX promotion section (sign-up credit of $0.25 to $0.75, GPT at a 0.1x rate, redemption codes in the comments), about 1,700 views each, with replies basically people leaving their ID to claim money; both sides of an invite get 8% on the first top-up, with cash commissions for promotion partners; checking sites such as BazaarLink and TokenAPI Scan list it automatically, and some checking sites have also issued fail reports questioning whether the models are genuine.

Inference: The 117,200 total visits are estimated from Similarweb global rank #290895 (fitted on 437 sites with real traffic, half of them within 17% error), so both channel circles carry ≈, and the channel shares and ranking are real. How each channel splits is inference: direct 78,500 = existing users 54,000 + invites and promotion partners 14,000 + V2EX posts 8,000 + checking sites and search 2,500; other channels 38,700 = referral links from invites and promotion partners 15,000 + checking sites and search 11,700 + V2EX posts 6,000 + existing users searching the brand name 6,000. The profile once recorded referrals 24.7% and search 4.9% (a different basis from the Similarweb public page), used only to judge that referrals are the bulk of the second circle. The 30% who come to register after being invited, 10% after seeing a V2EX post and 5% who click through after seeing it on a checking site are guesses; what veridrop.org is and how it relates to the site was not found.

Checkout visits, last month
8,596
#347 of 606
vs prior month
▼ 5%
9,010 → 8,596, gain −414
Website visits, last month
117,204
estimated from global rank; about 7.3 of every 100 visits reach checkout
Similarweb global rank
290,895
Take-off year
2026
How it startedExisting audienceThe maker brought an existing audience, community or client base into the product
New-user source nowOtherMostly referrals, email or paid, or mostly direct with no clear source
Growth engineReferral rewards or affiliatesRewards for referrals, or affiliate commissions
Spreads purely by itselfNo

Checkout visits over time

04–6 months ago avg
03 months ago
9K2 months ago
9KLast month

Checkout: checkout.stripe.com

1. Productwhat it sells, to whom, and how it charges

Product
Low-cost LLM API relay: one endpoint to call GPT, Claude, Grok, Gemini and image/video models, with per-request usage and cost logging.
Users
Developers who need cheap access to overseas LLMs, with Chinese users the largest group
Positioning
API relay cheaper than the official providers, with group-based failover, plugins and affiliate commissions
Pricing
Usage-based; about 200 models priced per million tokens/second/request/image; Claude-specific pricing requires a single top-up of US$50, OpenAI requires US$20 Billing: per item Free tier: no Source: Existing
Prices on site
usage-based USD per 1M tokens / per second / per request / per image, ~200 models (e.g. claude-opus-5-5 claude-premium USD 2, gpt-5.5 openai-stable USD 0.4, deepseek-v4-pro deepseek-stable USD 0.77) (credits)
Acquisition entry points on site
affiliate / referral program
Sharing mechanics
public profiles / leaderboards Project Plaza: 'Explore projects supported by or partnering with Modelflare.'; affiliate: 'Introduce your audience to Modelflare'
Price source
modelflare.dev
“Project Plaza: 'Explore projects supported by or partnering with Modelflare.'; affiliate: 'Introduce your audience to Modelflare'”Quote from the site

2. Channelswhere new users come from and what the traffic looks like

New-user source: Other (Direct / referral / email / paid). Direct 67%, referral 24.7% (veridrop.org 3.07%, game/account trading site z2u.com 0.78%, v2ex.com 0.7%), search only 4.9% (brand keywords 68%, plus model terms like grok api). New users come mainly from developer community referrals and affiliate promotion; existing users come back directly to top up and make calls.

Last month's traffic sources

Search 5%Referrals/email/paid 25%Direct 67%Unknown 3%

Top channel measured by Similarweb, others estimated by rank

Five-class attribution

Attributable new users, last monthShare
SEO4%
Other96%

Direct + brand search 68%. Direct and brand search count as viral only when the site shows sharing mechanics; the largest share wins, under 5% means Other.

Countries

CountryShare of visits
China43.91%
United States34.69%
Hong Kong9.27%
Singapore6.26%
Japan3.75%

Language

English

Keywords

78 keywords · brand ~68% · organic 86.69% / paid 13.31%

Similarweb top 5 keywords
modelflare
grok api
deepseek4.0 version-exp
qwen 3.8 mgsm
gpt6 vs fable

Non-brand queries: grok api, deepseek4.0 version-exp, qwen 3.8 mgsm, gpt6 vs fable

Referring sites

11 referring sites

SiteShare of referrals
veridrop.org3.07%
z2u.com0.78%
v2ex.com0.7%

Referrer categories

CategoryShare of referrals
Programming and Developer Software35.95%
Video Games Consoles and Accessories0.78%
Computers Electronics and Technology - Other0.19%
Computer Security0.17%

Site metrics (Similarweb, last month)

Visits MoM
+1.32%
Country rank
China #19144
Bounce rate
24.96%
Pages / visit
7.05
Visit duration
00:08:31
Industry
Unknown
Top 3 channels
Direct › Referrals › Organic Search (top channel 67.01%)

Growth-path traffic modelHow many people passed each step from launch to take-off

🔒 Unlock with quietmoney ProSee how many people passed each step from launch to take-off

3. Growth pathhow the first users arrived and what kept it growing

Classification

Basics

Timeline

Growth loop

Revenue and scale clues

Confidence and gaps

🔒 Unlock with quietmoney ProSee the timeline, growth loop, revenue clues and every source

Other sites that started via “Existing audience”

Similar sitessame category, also growing via Other

FAQ

How many checkout visits does Modelflare get per month?

About 8,596 last month (Similarweb estimate), #347 of 606; the month before was 9,010. Opening checkout is not paying, so this is not revenue.

Where do Modelflare's new users come from?

New-user source: Other (Direct / referral / email / paid). Mostly referrals, email or paid, or mostly direct with no clear source Direct 67%, referral 24.7% (veridrop.org 3.07%, game/account trading site z2u.com 0.78%, v2ex.com 0.7%), search only 4.9% (brand keywords 68%, plus model terms like grok api). New users come mainly from developer community referrals and affiliate promotion; existing users come back directly to top up and make calls.

How much does Modelflare cost?

Usage-based; about 200 models priced per million tokens/second/request/image; Claude-specific pricing requires a single top-up of US$50, OpenAI requires US$20

Dossier confidence: medium. Parts 1 and 2 come from reading the site and Similarweb data; part 3 from public reporting, founder posts, Wayback snapshots, WHOIS and similar, with speculation marked.

The number: checkout visits is Similarweb's estimate of visits from the site to a checkout payment page; opening checkout is not paying, so this is not revenue. Small sites land on fixed buckets.

Data through last month. Page updated 2026-10-31.