Tool-native
Trained on hand-written tool traces, not scraped chat. It learns the shape of a call, so new tools appended at inference route correctly without a retrain.
Homura-30B · open weights, Apache-2.0Tools for agents, and the model that calls them.
Download Homura and run it on hardware you already own. Point it at HYRE's paid endpoints. Pay per call in USDC — no key, no seat, no subscription.
Apache-2.0GGUF · MLX11 trained toolsNo API key
35 tools · priced per callPayment is the only gate.
The endpoint quotes a price, your agent settles it in USDC, the data comes back with a receipt. There is no account in the middle, because there is nothing to sign up to.
x402 · MPP · r4024 chainsQuote before executeFrom $0.001
The market for agent tooling.
Supply is the tools HYRE settles. Demand is the calls agents send. Price is set per call and quoted before execution, so an agent can decide whether the answer is worth the money before it spends any.
Model distribution
3 public builds · Apache-2.0
Counted103,643resolve requests
Hugging Face API · Homura-30B since 2026-08-13
Downloads counted103,643
Hugging Face's figure, which counts range-probes as well as real pulls. Read it with the arithmetic below, not on its own.
Priced tools35
18 HYRE endpoints and 17 partner services, across 13 families.
Cheapest call$0.001
The floor on the tools table. Priced per call, quoted before anything executes.
Settlement rails4
Solana, Base, BSC and Robinhood Chain. One balance behind them, and the chain only picks the stablecoin.
Priced right now
Live35 tools across 13 families, read from the production catalogue.
Quick Scan · Token Intelligence
$0.004
HYRE
Full Verdict · Token Intelligence
$0.015
HYRE
Whale Watch · Token Intelligence
$0.003
HYRE
Launch Forensics · Token Intelligence
$0.010
HYRE
Top Yields · Yield Finder
$0.002
HYRE
Migration Advisor · Yield Finder
$0.008
HYRE
Built from open weights
3 upstreamHomura is a thin LoRA on somebody else's frontier model. We say which one, and we say which part of the work was not ours.
Muse Glimmer 30BHomura v1 base
Qwen3.8-27BHomura v2 baseMuse-Glimmer-30B-hereticnot oursHomura builds
3 public reposEvery build is Apache-2.0 and free to download. Counts read from the Hugging Face API as this page renders.
Homura-30B103,476 downloadsHomura-Qwen3.8-27B117 downloadsHomura-30B MLX50 downloadshyrelabs on Hugging FaceReading the download number
the arithmeticEvery page that quotes a Hugging Face download count is quoting resolve requests. We would rather do the division in public than let someone else do it for us.
Hugging Face counts103,643
Across 3 builds, read from the API as this page rendered.
Which would be1,752 TB
Served, if every count were a full pull of a 16.9 GB file. It was not.
Downloads per like3,449:1
A genuinely popular GGUF repo runs 50–200:1. This is an order of magnitude out.
LM Studio, Ollama, Jan and mirror bots each issue header range-probes before pulling weights, and Hugging Face counts every one of them the same as a download. So the honest reading is that real human users are in the low thousands, not the tens of thousands — which is still roughly a hundred times the reach our gateway has ever had, and the only demand-side asset we own.
Run it from
5 runtimesGGUF and 4-bit MLX, so it loads in whatever you already have open. Nothing leaves the machine unless it is a paid call you chose.
OllamashippedLM Studioshippedllama.cppshipped
MLXshippedTransformersSettles on
4 railsOne balance, any endpoint. The chain only decides which stablecoin the call is priced in.
SolanaUSDC
BaseUSDC
BNB ChainUSD1 · USDT · USDC
Robinhood ChainUSDGWhat is Homura
The model is not the product.The tool call is.
Homura is the model HYRE built to drive its own rails. It is not tuned to chat with people — it is tuned to turn intent into the right paid call, then get out of the way. Open weights, Apache-2.0, running on hardware you already own.
It does not refuse
A refusal mid-chain is a system failure — there is nobody at 3am to click retry. On the internal bench Homura refused 0 of 7 legitimate technical prompts. The stock base model refused 2.
Local-first
Ships as GGUF and 4-bit MLX. LM Studio, Ollama, or bare llama.cpp on a Mac. No key, no vendor, no request leaving your machine unless it is a paid call you chose.
Run it on hardware you already own.
Download the weights, load them in LM Studio or Ollama, and the agent loop is yours. 560 hand-written training examples, 527 tool calls, zero malformed JSON — and the whole tune cost $12.10.
Two formats
16.9 GB GGUF for llama.cpp and Ollama, 21.3 GB 4-bit MLX for Apple Silicon. Both quantised to fit a 32-64 GB machine.
Apache-2.0
Commercial use allowed, modification allowed, attribution required. No gated repo, no waitlist, no acceptance form.
Vision intact
The tune touches the language tower only — all 832 adapted tensors. The vision tower was never modified and still scores 6/6 on the internal check.
- What we changed
- A LoRA adapter of about 200 MB on top of a 60 GB model, trained on tool traces we wrote by hand. The base model is Meta's and stays Meta's.
- What we did not
- The abliteration. That is community work by darkc0de, credited on the model card, and we did not redo it — it saved GPU time and the risk of breaking other capabilities.
- What it cost
- About $12.10 of GPU time end to end, infrastructure mistakes included. Two working days from research to a published repo.
One protocol, every tool on the rails.
11 tools are in the trained surface — resolve_token, get_token_price, swap_quote, bridge_quote and the rest. Anything beyond them is appended to the system prompt at inference and the model still routes it, because it learned the pattern rather than the names.
35 priced tools
Trench data, token prices, whale flow, wallet PnL, Meteora DLMM pools, Nansen smart money and cross-chain yield — 13 families, each call priced on its own.
MCP server
Install once in Claude Desktop or Cursor and the same tools show up in your editor, settled the same way.
Discovery
bazaar_search finds an x402 endpoint the model has never seen, quotes it, then pays it. Quote first, execute second — always.
- Trained tools
- 11 tools are in the fine-tune itself, drilled through 527 hand-written calls with no malformed JSON in the set.
- Untrained tools
- Appended to the system prompt at inference. The model was never shown them and still selects them correctly, because the training taught the shape of a call rather than a list of names.
- Discipline
- Quote, then execute. The model is drilled to price a call before making it, which is the only safe order when the call spends money.
Payment is the only gate.
There is no API key to issue, rotate or leak. The agent presents payment, the endpoint answers. That is the whole authorisation model, and it is the reason an agent can use HYRE with no human in the loop.
Four settlement rails
Solana through x402/MPP, Base through PayAI, BSC through b402, and Robinhood Chain through r402 — the first x402 rail on that chain, where calls settle in USDG with no gas.
Priced per call
USDC, quoted before execution and capped server-side. No subscription, no minimum, no seat.
The model holds nothing
No private key, no send authority. Spend limits live in the payment layer, where a limit can actually be enforced.
- What you pay
- The per-call price the endpoint quotes, in USDC. It is on the tools table below, and it is the same number the backend serves to the production playground.
- What holds the key
- Not the model. The weights have no wallet and no send authority; spend limits sit in the payment layer, which is the only place a limit can be enforced.
- What the agent buys
- Inference itself. When Homura cannot serve a call, HYRE buys the completion from the UsePod marketplace over x402 — quoted, capped, then settled in USDC on Solana from the caller's own wallet, one call at a time.
- What a chain changes
- Only which stablecoin settles. Solana and Base take USDC, BSC also accepts USD1 and USDT summed server-side, and Robinhood Chain settles in USDG gaslessly.
Receipts, not vibes.
Every settled call leaves an on-chain receipt, every agent can carry a portable identity, and a share of revenue buys $HYRE back and burns it. All three are verifiable without asking us.
Settlement receipts
Each paid call writes a receipt on its settlement chain. Read it yourself — no dashboard required.
Portable identity
Agents carry a .ME passport and ERC-8004 registry entry, so the thing calling your endpoint can be identified across runs.
Buyback and burn
Ecosystem revenue buys $HYRE from the open market and destroys it. Every burn is verifiable on-chain.
- The call
- Settles on its own chain and leaves a receipt there. You do not have to trust a dashboard for it, and neither does an agent.
- The model
- Download counts and likes come from the Hugging Face API, which is the same number anyone else sees on the repo page.
- The token
- Buybacks and burns happen on-chain. If a burn is claimed and not on the chain, it did not happen.
How a call gets routed
Local costs nothing, so buying is a decision.
Homura runs on hardware we own, so a call it can handle has no marginal cost. That turns frontier compute into something the agent buys on purpose — from the UsePod marketplace, over x402, one call at a time. Move the controls and the router below answers exactly as the backend would — same policy, same reasons.
Which model did the caller ask for?
How big is the prompt?1,200 / 24,576 tok
Is the local node awake?
Answers locallylocal-defaultcosts nothing
Nothing forced a purchase, so the model we host answers and the call costs nothing.
What a UsePod call costs, right now
Reading—
A real 402 from UsePod for a model at — max tokens. Quoting is free, so nothing was paid to show you this.
- Pay to
- —
- Network
- —
- Mode
- —
- Quote expires
- —
What a settled call leaves behind
- signature
- The Solana transaction that paid for this call
- capMicroUsdc
- What we were willing to pay on-chain, in USDC microunits
- route
- Which UsePod operator, or the centralized fallback, served it
- providerId
- Which operator answered
- gatewayReceipt
- The gateway's own countersigned record of the charge
These are the fields, not a sample transaction. Signatures appear here once calls settle on mainnet — we are not going to print one that never happened.
The tools, live
One row per call an agent can pay for.
18 HYRE endpoints and 17 partner services, read from the production catalogue when this page rendered. The price column is what settles.
| Tool | Family | Returns | Served by | Rail | Per call |
|---|---|---|---|---|---|
| Quick Scantoken-intel:snipers | Token Intelligence | Detect sniper wallets that bought at launch | HYRE | $0.004 | |
| Full Verdicttoken-intel:verdict | Token Intelligence | Holders + snipers + bonding curve + risk assessment | HYRE | $0.015 | |
| Whale Watchtoken-intel:whales | Token Intelligence | Top holders with % of supply | HYRE | $0.003 | |
| Launch Forensicstoken-intel:forensics | Token Intelligence | First-60s birth-window rug-risk score (bundle/dev-dump/concentration) + clean|suspect|rigged band | HYRE | $0.010 | |
| Top Yieldsyield-finder:quick | Yield Finder | Top yield pools sorted by APY | HYRE | $0.002 | |
| Migration Advisoryield-finder:advisor | Yield Finder | Cross-chain yield comparison + break-even math | HYRE | $0.008 | |
| PnL Snapshotwallet-analyzer:pnl | Wallet Analyzer | Realized + unrealized PnL, portfolio value | HYRE | $0.005 | |
| Full Intelwallet-analyzer:intel | Wallet Analyzer | 30-day trading patterns + success rate + archetype | HYRE | $0.012 | |
| Pool Listlp-analytics:list | LP Analytics | Top DLMM pools with APR/TVL/volume | HYRE | $0.001 | |
| AI Recommendlp-analytics:recommend | LP Analytics | Best pool + entry strategy (spot/curve/bidask) | HYRE | $0.008 | |
| New Launchesnew-tokens:new | Trench Watch | Latest PumpFun token launches with snipe signal | HYRE | $0.008 | |
| Graduatingnew-tokens:graduating | Trench Watch | Tokens near bonding curve graduation (>70% progress) | HYRE | $0.003 | |
| Chain TVLdefi-overview:tvl | DeFi Overview | Total Value Locked across DeFi chains | HYRE | $0.001 | |
| Ask DeFi Expertask:default | Ask Anything | AI-orchestrated multi-source DeFi answer | HYRE | $0.025 | |
| Smart Money Flownansen:flow-intel | Nansen Smart Money | Net-flow breakdown: Smart Traders vs Whales vs Public Figures for one token | HYRE | $0.020 | |
| Token Screenernansen:screener | Nansen Smart Money | Find tokens that smart money is actively accumulating | HYRE | $0.025 | |
| PnL Leaderboardnansen:leaderboard | Nansen Smart Money | Top wallets ranked by realized PnL on one specific token | HYRE | $0.030 | |
| Wallet Deep Intelnansen:wallet-intel | Nansen Smart Money | Full wallet profile: archetype, win rate, top tokens, recent smart trades | HYRE | $0.050 | |
 Nansen Perp Screener (raw)external:nansen-perp-screener-raw | External Services | Perpetual contract screening data — direct from Nansen, no enrichment | api.nansen.ai | $0.025 | |
| Nansen Perp Leaderboard (raw)external:nansen-perp-leaderboard-raw | External Services | Top perpetual traders by PnL — direct from Nansen, no enrichment | api.nansen.ai | $0.075 | |
| Xona AI Video Generationexternal:xona-video-gen | External Services | Generate a short AI video from a text prompt (Grok Imagine Video). Returns video URL. | api.xona-agent.com | $0.600 | |
| Amadeus Flight Searchagentcash-travel:amadeus-flights | Travel | Amadeus GDS flight offers — real bookable fares by route and date | stabletravel.dev | $0.110 | |
| Google Flights Searchagentcash-travel:google-flights | Travel | Google Flights offers + price insights for a route and date | stabletravel.dev | $0.040 | |
| Hotel Searchagentcash-travel:hotel-search | Travel | Hotels and villas by location, dates, guest count, and budget | orbonomy-xyz.vercel.app | $0.100 | |
| Exa Neural Searchagentcash-web:exa-search | Web Search | Neural web search — semantic results for research-style queries | stableenrich.dev | $0.020 | |
| Exa AI Answeragentcash-web:exa-answer | Web Search | Direct AI-generated answer with source citations | stableenrich.dev | $0.020 | |
| Google Searchagentcash-web:serper-search | Web Search | Classic Google Search results via Serper | stableenrich.dev | $0.040 | |
| Firecrawl Scrapeagentcash-web:firecrawl-scrape | Web Search | Scrape a single URL with JS rendering — returns page content | stableenrich.dev | $0.025 | |
| Google Newsagentcash-web:serper-news | Web Search | Google News results via Serper — latest headlines for a query | stableenrich.dev | $0.080 | |
| Place Text Searchagentcash-maps:text-search | Maps & Places | Free-text Google Maps search — places matching a query string | stableenrich.dev | $0.040 | |
| Nearby Searchagentcash-maps:nearby-search | Maps & Places | Places around a lat/lng point, filtered by type and radius | stableenrich.dev | $0.040 | |
| Place Detailsagentcash-maps:place-details | Maps & Places | Full details for one place — hours, rating, contact (needs placeId) | stableenrich.dev | $0.040 | |
| Apollo People Searchagentcash-people:apollo-people-search | People & Company | B2B prospect search by title, company, and location filters | stableenrich.dev | $0.040 | |
| Apollo People Enrichagentcash-people:apollo-people-enrich | People & Company | Enrich one person by email, name, or company domain | stableenrich.dev | $0.100 | |
| Minerva Person Enrichagentcash-people:minerva-enrich | People & Company | Person enrichment by PID, LinkedIn URL, email, or phone | stableenrich.dev | $0.100 |
Questions
The ones we actually get asked.
What exactly is Homura?
A 30B open-weight model, fine-tuned by HYRE to call HYRE's tools. The base is Meta's Muse Glimmer 30B under Apache-2.0, run through a community abliteration, then given a thin LoRA trained on hand-written tool traces. We did not pretrain a model, and we say so on the model card.
Why an uncensored model?
Because for an agent a refusal is a bug, not a safety feature. An automated chain that stops halfway because the model declined to name a number has failed, and there is no human there to retry. The real limit is that the model holds no key and cannot move funds — the brake is in the payment layer, not the personality.
What did HYRE actually contribute?
The LoRA and the dataset. 560 conversations written by hand — roughly 70% tool-call traces and 30% house voice — producing 527 tool calls with zero malformed JSON. Training loss went 2.4 to 0.048 over three epochs, for about $12.10 of GPU time.
Do I have to use Homura to use the tools?
No. Every tool is plain paid HTTP, so any model or script that can present payment can call it. Homura is what we run ourselves, and what we recommend if you want the whole loop to stay on your machine.
How does paying work without an account?
The endpoint answers a bare request with a 402 and a price. Your agent settles it on whichever of the four chains it holds a balance on — USDC on Solana or Base, USD1/USDT/USDC on BSC, USDG gaslessly on Robinhood Chain — then retries and gets the data. There is nothing to sign up for, because the payment is the authentication.
Where do I verify any of this?
Weights and download counts on Hugging Face, settlement receipts on the chain the call settled on, burns on-chain. The numbers on this page come from those sources and nowhere else.