How to run a FLOP GPU miner on devnet: step-by-step guide

Update, November 9, 2026: Flop's miner manual, now marked updated November 3, says "from 2 November 2026 no invitation is needed". The sponsor approves new devnet miners automatically within a cap, and a public release pack replaces the operator's access bundle. This guide now follows that version. Got a home GPU? See mining FLOP on a home GPU.

Updated November 9, 2026. The FLOP devnet opened for builders on November 3. This guide follows Flop's miner manual, last updated November 3, as read on November 9. We quote its commands and add none of our own. The release pack's README comes first if it differs from this page.

Devnet earns nothing. Flop's timeline FAQ says "No rewards, no airdrop weight, no scoring." The miner manual adds that it "does not promise mainnet participation or earnings." Devnet FLOP has no value and the chain may be reset. FLOP never asks for your seed phrase. Check every link against flop.finance/official or our official links list.

Can anyone run a FLOP miner on devnet?

Yes. The manual says that from October 30, 2026 "miners join the devnet over public endpoints", and that "from 2 November 2026 no invitation is needed: the sponsor approves a new devnet miner automatically while at most 100 automatically sponsored miners are still uncalibrated, the new one included." Past that cap the sponsor answers 429, and you retry once one of the pending miners calibrates.

Running several miners? The manual says to "request operator approval once instead", giving your operator key's address and planned miner count. FLOP then lists your key with a quota, and you approve each new miner yourself with the pack's sponsor-grant.sh. A grant lasts 30 days by default.

What works today, per the manual: "sponsorship, registration and calibration through the public mailbox. Paid sessions are not open yet." So a devnet miner can get registered and calibrated, but it can't serve paid inference yet.

If you're new to this, Flop's plain-language page Put your GPU to work on FLOP is the gentler start. For a wider look at hardware and testnet plans, see how to mine FLOP and our miner setup guide.

How does a FLOP miner work?

The manual describes two services: "a CPU miner-worker paired with a GPU miner-vllm-toploc backend. The backend runs the model; the worker signs chain operations, handles compute-channel sessions, publishes TOPLOC evidence, and renews host-wide calibration."

SOFT is the default tier and "needs no confidential-computing hardware". The manual also says the person who runs the host owns the miner account: "Renting out GPU capacity alone does not make anyone a miner."

What GPU and host do I need?

"The validated serving drill used an NVIDIA A40 with Qwen2.5-0.5B-Instruct, BF16, and the worker and backend on the same host." Other NVIDIA cards may work if they fit the model, each with its own throughput check. Flop's testnet page lists a provisional miner profile of a single GPU or cluster with 16 GB+ VRAM per unit.

Before you pay for a long lease, the manual says to check:

• A full VM or owned host with Docker, Compose, the NVIDIA driver and Container Toolkit. "A provider's container-only product may forbid nested Docker."
• The GPU shows up in nvidia-smi on the host and in a Docker GPU test.
• Enough disk for the exact checkpoint, with persistent env, seed and worker volume.
• A public DNS name with ports 80 and 443 open to the host, or your own tunnel.
• A way to delete the instance and read your bill. "Stopping Compose does not stop VM billing."

Which GPU rentals does the manual cover?

The FAQ section below covers GPU.ai and RunPod. In short, a GPU.ai VM is "the documented rental for same-host Compose". A RunPod GPU Pod can only run the backend, with the worker on a separate Docker host.

What are the devnet limits?

From the manual's "Devnet limits" table, as read on November 8, 2026, after Flop Labs raised the limits 10x on November 7:

Limit	Value
New sponsored miners	At most 100 per rolling 24 hours (was 10)
Sponsored stake	At most 200 million FLOP per rolling 24 hours (was 20 million), new stake and top-ups together. A new miner starts at about 1 million FLOP; top-ups raise it to at most 2 million
Sponsorships per account	One
Public sponsor route	6 requests per minute per address

"Requests past a daily limit get 429 and nothing is registered." Check the live budgets at https://faucet.flop.network/limits before you ask. Our FLOP devnet limits page explains each number.

Step 1: Download the public release pack

The manual's release pack was published at https://flop.network/releases/cli/miner-pack-devnet-ts-20261106/. It gives these commands to fetch it and check its checksum:

1. B=https://flop.network/releases/cli/miner-pack-devnet-ts-20261106
2. curl -fLO "$B/flop-miner-pack-devnet-ts-20261106.tar.gz" && curl -fLO "$B/README.md"
3. echo "35f222f748447a8aabdd015d94260d8facfd29305691c3cf7ed033cbb2b514b1 flop-miner-pack-devnet-ts-20261106.tar.gz" | sha256sum -c -

Stop if the checksum doesn't match. Per the manual, the pack supplies:

• The public endpoints: FLOP_RPC is wss://rpc.devnet.flop.network, GHOST_MAILBOX_URL is https://mailbox.devnet.flop.network, and the stake sponsor is https://faucet.flop.network (POST /sponsor/miner).
• A DA_URL value, which "the worker no longer uses" but Compose still requires.
• The worker, backend and CLI images with the Compose files and the env template they match. The images "pull anonymously" from ghcr.io/flop-labs/flop-core.
• The registered models in its releases.json, where you pick MODEL_HASH and the exact 40-character checkpoint revision.

"You send nothing to FLOP." The sponsor request carries the miner address, and the worker publishes your public DNS name in its signed endpoint ad.

Can a coding agent set up the miner for me?

The manual describes an agent path: a coding agent "can follow the public flop-miner-devnet skill with the devnet operator's pinned public installer pack." On an Ubuntu 24.04 NVIDIA GPU VM, you extract the pack into a fresh private directory and run:

1. tar xzf flop-miner-pack-public.tar.gz
2. cd flop-miner-pack
3. sudo ./install.sh miner-01 --public-host miner.example.com

Use your own machine and DNS names. The DNS name must resolve to the host, and inbound ports 80 and 443 must reach it. --install-deps consents to installing dependencies. The manual says this one-command installer pack "is not published yet; until it is, an agent follows the release pack's README instead." It adds that "sponsorship is automatic on devnet".

The skill file returned a 404 page when we checked on November 3. On November 9 it loaded. It asks for "a full Ubuntu 24.04 NVIDIA GPU VM or owned host, with at least 16 GB VRAM and 50 GiB model disk budget" and points to the same release pack as Step 1.

Step 2: Create the miner account

The manual asks for "a dedicated sr25519 miner account, separate from every validator account." Its preflight runs the flop CLI from its container image:

• FLOP_CLI=ghcr.io/flop-labs/flop-core/cli:stable
• flop() { docker run --rm --network host -e FLOP_KEY_PASSWORD -v /opt/flop-miner:/work "$FLOP_CLI" "$@"; }
• flop --version (the manual says the image prints 0.4.0)

Check the network before any signed transaction:

• flop --network devnet --rpc wss://rpc.devnet.flop.network network status --json
• nvidia-smi
• docker info

"Stop on mismatch." Generate the seed offline, write it as one line to a private file, back it up offline, then install it with sudo install -o root -g root -m 600 seed.txt /opt/flop-miner/signer. "Never use //Alice or a validator key."

Step 3: Get sponsored stake

From the miner host, the pack's helper signs a request with the miner key and asks the sponsor for stake:

sudo env SPONSOR_URL=https://faucet.flop.network FLOP_RPC=wss://rpc.devnet.flop.network ./sponsor.sh

"Do not register the account yourself first: the sponsorship registers it, at a 10% commission you can change later." The sponsor "admits a devnet miner automatically within the cap" and gives one sponsorship per account. It posts about 1 million FLOP of locked stake, which the miner can use only as stake, plus a small grant for fees.

A 429 means automatic sponsorship is full or a daily limit was reached. A 403 means the address is listed for the other role, or a grant is expired or not from an approved operator. A 503 means retry later. Check https://faucet.flop.network/limits before you ask: when we read it at 10:31 UTC on November 9, the miner sponsor status showed unknown with reason sponsor_read_timeout, so it's worth checking more than once.

Step 4: Fetch the model and decode policy

The manual fetches the checkpoint with the CLI, which writes files only if they match the on-chain root:

flop --network devnet --rpc wss://rpc.devnet.flop.network model fetch "$MODEL_HASH" --precision BF16 --output-dir /work/models/qwen2.5-0.5b-instruct --json

"Do not substitute a manual download." Nobody sends you the decode policy. Leave DECODE_POLICY and DECODE_POLICY_HASH empty in the env file, then run the pack's resolver before the first start:

sudo ./decode-policy.sh miner-01.env

Step 5: Fill in the env file

Make a private miner-01.env from the template in the release pack, one copy per miner. The manual's key values include:

• FLOP_RPC=wss://rpc.devnet.flop.network
• GHOST_MAILBOX_URL=https://mailbox.devnet.flop.network
• OPENAI_BASE_URL=http://backend:8000/v1 for the same-host setup
• SESSION_BIND_IP=127.0.0.1 and SESSION_HOST_PORT=8645
• SESSION_PUBLIC_HOST set to your public DNS name, lowercase, without https://
• INFERENCE_API_KEY and SESSION_AUTH_TOKEN, each a fresh openssl rand -hex 32 value

The manual also has you create a session-signing seed for MINER_ENCLAVE_KEY_FILE before the first start, and says to back it up: "a new seed means a new session key."

Step 6: Start the backend, then the worker

The manual defines a compose shell function over its three Compose files (leave out the TLS file if you run your own tunnel). Then:

1. compose up -d backend
2. Check the backend's private /health and confirm model_identity: "measured", the expected model hash, precision and decode-policy hash.
3. compose up -d worker tls

From a machine outside your network, a request without the token should answer 401. The manual's check is curl -s -o /dev/null -w '%{http_code}\n' https://<SESSION_PUBLIC_HOST>/session/info.

Step 7: Calibration

After registration, Flop's issuer sends calibration jobs. "From then on keep the worker running: a job left unanswered past its 600-block deadline expires and slashes the miner." The manual says the source requirement is 64 verified jobs over at least 600 finalized blocks, and "a deployed devnet may still require 512 jobs over 21,600 blocks." The cap expires after seven days unless renewed. "An uncalibrated host is not ready: /session/open returns 503."

How do I shut a miner down?

"To retire a rental, stop Compose, delete the provider instance, and confirm in the provider console that it is gone and billing stopped. When any host retires, ask the operator to remove its enrollment." Avoid --volumes unless you mean to delete calibration state.

FAQ

Is FLOP devnet mining open to everyone?

Yes. The miner manual says that from November 2, 2026 no invitation is needed. The sponsor approves new devnet miners automatically as long as no more than 100 automatically sponsored miners, the new one included, are still uncalibrated. Paid sessions are not open yet.

Does devnet mining earn FLOP tokens or airdrop points?

No. Flop's timeline FAQ says the devnet earns nothing, with no rewards, airdrop weight or scoring. There is no FLOP token yet, and paid devnet sessions were not open on November 3, 2026.

What GPU does the FLOP devnet miner need?

Flop validated its drill on an NVIDIA A40 running Qwen2.5-0.5B-Instruct in BF16. Other NVIDIA cards may work if they fit the model, but each needs its own throughput and TOPLOC check.

Can I run the FLOP miner on RunPod?

Only as the backend. A RunPod GPU Pod can't run the Compose pair, so the worker runs on a separate Docker host. The manual lists 60 GB of container disk for Qwen2.5-0.5B and port 8000. RunPod serverless endpoints are unsupported.

Can I run the FLOP miner on GPU.ai?

A GPU.ai VM is the documented rental for running both services on one host. GPU.ai documents no public inbound port, so you need your own tunnel. Flop hasn't verified whether every GPU.ai VM comes with Docker and the NVIDIA runtime.

What happens if my devnet miner goes offline?

A calibration job left unanswered past its 600-block deadline expires and slashes the miner, so keep the worker running once calibration starts.