HyperFusion: three open-weight models, and where the GPUs sit

HyperFusion at hyperfusion.x402.press serves three open-weight models over x402 at the same per-call price as the DeepSeek endpoint for the same model, on GPUs domiciled in the United Arab Emirates.

What it is

HyperFusion is a hosted endpoint at https://hyperfusion.x402.press. It answers POST /v1/chat/completions in the OpenAI shape, so an agent built against an OpenAI client needs its base URL changed and a payment step added.

There is no sign-up, no API key and no monthly bill. A call is paid for with an x402 payment against the HTTP 402 challenge the endpoint returns; the completion comes back on the retried request that carries it. The service holds its own credentials with the upstream provider, server-side, so there are no keys to issue or rotate.

The models are those already served elsewhere on this network, at the same per-call price; the difference is where the inference runs. The service states that inference runs on GPUs domiciled in the United Arab Emirates, and that prompts and completions stay in-country. We report that as the service makes it, and did not audit it.

For a buyer whose requirement is that inference runs on UAE-domiciled hardware, that statement is the reason to choose it. For everyone else it is the same inference at the same price with one feature missing.

What you get

Three models

Table 1. The three models served, context window as stated on the service card, price per call on Base. Source: GET /v1/models and the live 402 quote, 30 October 2026.

The upto amount is the ceiling at max_tokens: 1000; it rises with the output allowance you authorise (see below).

The identifiers in Table 1 are the full names GET /v1/models returns, and they matter. deepseek-v4-flash, the short id used by the DeepSeek endpoint on this network, is rejected with HTTP 400 and unsupported_model; porting between the two means changing the model string and nothing else.

The request

The body is the OpenAI chat completions body. model and messages are the parts you always send.

{
  "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
  "messages": [
    {"role": "system", "content": "Classify each support ticket."},
    {"role": "user", "content": "My invoice shows a charge I do not recognise."}
  ],
  "temperature": 0.2,
  "max_tokens": 200
}

The model is required: a body without one returns HTTP 400 with model_required, and there is no default, so no call is billed to a model the client did not name. temperature, max_tokens, top_p, frequency_penalty, presence_penalty and stop are forwarded, so tuned sampling behaviour carries over.

One parameter is refused. "stream": true returns HTTP 400 with stream_unsupported, and the reply arrives whole.

The free calls

GET /v1/models lists what the endpoint serves. GET /hyperfusion/feedback returns the 50 most recent entries; POST to the same path submits one against a response_id, with a rating from 1 to 5 and an optional comment. /health, /about and /llms.txt are readable. None is billed.

Feedback is the only evidence of model quality available here: no benchmark is published. It is a record, not a support line — no ticket, no reply timescale.

What it costs

Six chains quote two amounts for a call; Avalanche quotes exact only. Under exact you pay a flat prepaid amount; under upto you authorise a metered ceiling and the charge settles on measured tokens, never above it. The price is set by the model named, by which of two prompt bands the request falls into — split at roughly 4,000 estimated tokens, as the card puts it — and, under upto, by the max_tokens you authorise: the ceiling is the most the call can cost at that allowance. There is no peak-hour surcharge.

Table 2. The two models priced at $0.00700 on Base: price per POST /v1/chat/completions in US dollars, by settlement chain and payment scheme. Source: live 402 handshake, 30 October 2026, outside peak windows.

The upto amount is the ceiling at max_tokens: 1000; it rises with the output allowance you authorise (see below).

Table 3. MiniMaxAI/MiniMax-M2.7: the same call on the third model. Same source.

The upto amount is the ceiling at max_tokens: 1000; it rises with the output allowance you authorise (see below).

The chain is the larger lever (Table 2). On Base a flash-tier call costs $0.00700 under exact; on Avalanche the same call costs $0.00200, 71% less. On Ethereum mainnet it costs $3.00 — 428 times the Base price, and the same multiple under upto at a 1,000-token ceiling: $4.092858 against $0.00955. Nothing about the completion changes. Arbitrum at $0.00600 and Polygon at $0.00500 sit between. Base is the sensible default: it quotes both EURC and USDC.

The model is the smaller lever, and a flat one (Table 3). Under exact, MiniMax costs 1.89 times a flash-tier call on all seven chains, because the two price lists move together. Under upto at a 1,000-token ceiling the gap widens to 2.72 times, since the ceiling sits 36% above the exact amount on the flash tier and 96% above it on MiniMax. The flash-tier models are quoted identically everywhere, so choosing between DeepSeek and Qwen is not a pricing decision.

Treat the tables as reference points, not quotes: the 402 response carries the amount you pay. The service’s llms.txt lists $0.00875 on Base for the DeepSeek tier, which is not what the endpoint quotes. We use the quote.

What the ceiling actually does

The upto amount is not a second price. It is the most the call can cost at the max_tokens you authorise, so it moves with that allowance rather than with the work the call does. Table 4 measures it on Base with one short prompt, changing nothing but max_tokens.

Table 4. DeepSeek-V4-Flash-0731 and MiniMaxAI/MiniMax-M2.7 on Base: the upto ceiling at four output allowances, one short prompt, exact unchanged throughout. Qwen/Qwen3.8-27B is quoted identically to DeepSeek on every chain in Table 2 but was not measured at other allowances. Source: measured 30 October 2026.

The flash-tier ceiling is flat from 1,000 to 8,000 and moves only above it. Set max_tokens deliberately: a call that leaves the field out authorises the model’s maximum output, and the flash ceiling rises from $0.00955 at 1,000 tokens to $0.129261 with the field unset — about thirteen and a half times the figure in Table 2 — on a prompt that has not changed. The ceiling is a limit on what a call can cost you, not a forecast of what it will cost.

How an agent uses it

Illustrative worked example. A Dubai-based operations agent makes 500 calls a day: 90% to DeepSeek-V4-Flash-0731, 10% escalated to MiniMaxAI/MiniMax-M2.7, settled on Base under exact with short prompts, so $0.00700 and $0.01325 are the amounts it pays.

• Flash: 450 × $0.00700 = $3.15 a day
• MiniMax: 50 × $0.01325 = $0.6625 a day
• Total: $3.8125 a day, or $114.38 a month at 30 days

All-Flash at the same volume: 500 × $0.00700 = $3.50 a day, or $105.00 a month. The escalation adds $0.3125 a day, $9.38 a month, or 8.9%; ten per cent of the calls carry 17% of the bill. Settled at the upto ceiling at max_tokens: 1000 instead, the same workload would cost $5.595 a day, or $167.85 a month — the upper bound at that allowance. With max_tokens unset the ceilings are $0.129261 a call on the DeepSeek tier and $0.476301 on MiniMax, and that bound would not hold.

The chain moves the bill further than the routing does: the same 500 calls cost $1.08925 a day on Avalanche ($32.68 a month), $2.72320 on Polygon ($81.70), $3.26785 on Arbitrum ($98.04) and $1,633.93 on Ethereum mainnet ($49,017.86).

So what does residency cost? On the model carrying 90% of this traffic, nothing: $0.00700 a call here is $0.00700 a call on the DeepSeek endpoint, and the all-Flash month is $105.00 either way. The choice is not about price; it is about where the compute sits and which features you need, and this side lacks streaming. The $114.38 month is not a premium for residency either; it is the ordinary cost of routing one call in ten to a second model.

If the requirement is real, the comparison is not between two prices but between a data path you can use and one you cannot. If it is not, this is the same model at the same price with a feature removed. Either way, set the escalation trigger from what the task requires — a task class, or an output the agent can validate against its own schema — and check it against the feedback record. A retry after a failed cheap call pays twice.

Where it stops

Five limits matter before deployment.

No performance figures. The service publishes no latency, no throughput and no uptime figure. Its only operational numbers are a health flag at /health and a cumulative call counter at /stats, neither of which measures speed or reliability. If those properties decide usability, measure them against your own traffic.

No streaming on the paid route. "stream": true is refused with HTTP 400. For a chat interface this is a real constraint, not a footnote: nothing can be rendered until the completion arrives, and a client expecting server-sent events needs a different response path.

Three models only. DeepSeek V4 Flash, Qwen3.8-27B and MiniMax M2.7, and an unknown identifier is a 400. A workload needing a fourth model needs a second endpoint.

The prompt band is not fully documented. The card names two bands split at roughly 4,000 estimated tokens. We measured two prompt sizes only: a short prompt, which took the lower band, and one of about 48,000 characters, which took the upper band on every chain. We did not locate the boundary, so a prompt near the split cannot be priced from the card.

No service level and no agreement behind a call. There is no SLA, and no signed data-processing agreement. The residency statement is about where the GPUs are; it is not a contract, and we did not verify it. The service card labels its own maturity beta. Where a reviewer, a client or a regulator requires a signed instrument governing the data, none of this substitutes.

When not to use it

Three cases where a different instrument is the right one.

You need streaming. If the product is a chat interface, or first-token latency is part of the experience, this endpoint cannot serve it. No price advantage closes a capability gap.

You need a model outside the three. If the requirement is a larger model, a different family or a fine-tuned one, the endpoint is not a candidate, whatever it costs.

You need a signed agreement governing the data. A statement that inference runs on UAE-domiciled GPUs is a fact about hardware, not a contract. Where the requirement has to be evidenced by an instrument rather than a statement, this endpoint does not provide the evidence, and should not carry data that requires one.

For everyone else the decision is small and reversible: no commitment, no minimum, no account; a migration is a base URL and a model string.

How to start

Two requests, with one payment in between.

1. Call GET /v1/models. It is free, it returns the three identifiers, and it confirms your base URL and network path before any money is involved. Use the exact strings: deepseek-ai/DeepSeek-V4-Flash-0731, Qwen/Qwen3.8-27B, MiniMaxAI/MiniMax-M2.7.
2. Send POST /v1/chat/completions with a model and a messages array, and with stream set to false or omitted. The first attempt returns HTTP 402 with the amount due and the payment details for your chain. Six chains quote a metered ceiling alongside it; Avalanche quotes the exact amount only.
3. Pay the quoted amount in USDC, or in EURC on Base, from a wallet holding a small balance, then present the payment. The completion comes back on the retried request.

Pointing an existing OpenAI client at https://hyperfusion.x402.press/v1 and adding that payment step is the whole of the integration. The only state an agent needs is its float: $5 of stablecoin holds 714 calls at $0.00700, and about 655 at the 90/10 mix.

Take the amount from the 402 response and pay that. The tables are for budgeting; the handshake is for paying.

Method note

Prices read from the live 402 handshake on 30 October 2026, outside peak windows: three models, seven chains, both schemes where offered. Identifiers come from GET /v1/models. The refusal of a streaming request, a missing model and an unknown identifier were observed on the live endpoint and are quoted as returned. The service’s llms.txt quotes $0.00875 on Base for the DeepSeek tier; the live quote is $0.00700 and governs. The Dubai example is arithmetic, not a measurement. The upto ceilings were measured by varying max_tokens — 1,000, 8,000 and 32,000, then unset — on Base with the prompt held fixed, also on 30 October 2026. The residency statement is reported as the service makes it and was not independently verified.

Filed under Service Reviews · Tags: x402, inference, hyperfusion, data-residency

Related reviews: DeepSeek review · Qwen review · SpiralStake review · Wolfram Alpha review · ElevenLabs review · Mage review

Background: what x402 is, and how a payment works

All seven reviews: the services reviewed

Marketplace: HyperFusion on x402.press — prices per chain, operations and the metered ceiling.