L00P.AI / RASPBERRY PI 500+
An AI workspace.
On your SSD.
A free image bringing AI agents, the Szpieg+ workspace and your audio archive together. We are preparing it for leaders who want to explore the technology themselves and set the direction of its deployment.
We have a US-keyboard Pi 500+. We will build the workspace from scratch on that device, then prepare an image for torrent distribution. The image and test results are not available yet.
WHAT BRINGS IT TOGETHER
A folder of recordings becomes a starting point.
In the planned first-run workflow, you will choose an archive directory. Originals will remain read-only; transcripts, indexes and exports will go to a separate folder. Words will retain their recording timestamps, so important answers can be checked by listening.
Recording
The original and its provenance.
Words and time
Karaoke transcription and metadata.
Your workspace
Listening, retrieval and agent workflows.
Explore the existing interface demonstration. It shows how Szpieg+ works; readiness of the Raspberry Pi port will be assessed separately. Open the demo →
Choose where the model runs.
ON YOUR COMPUTER
Local audio and small models
Plan: Szpieg+, whisper.cpp / sherpa-onnx, Ollama or llama.cpp. After model download, this path is intended to work offline. Acceptance includes a disconnected test. Agents will receive access only to selected tools and directories.
WITH YOUR OWN ACCOUNT
Claude Code · ChatGPT / Codex
A client application does not include Claude or GPT model weights. Cloud services require Internet access, your own account and a suitable plan or API billing. Sending material to the cloud must be the user’s deliberate choice.
Hermes Agent and OpenClaw connect tools to a chosen model provider. Running an agent on a Pi does not mean a large model runs locally. OpenClaw’s Raspberry Pi guidance recommends cloud models. OpenClaw / Raspberry Pi.
ARM64: what is documented and what still needs testing?
Claude Code, Codex CLI, Hermes and Ollama have Linux ARM64 installation paths. ChatGPT now offers a Linux ARM64 preview, and LM Studio documents Linux ARM64 requirements. Raspberry Pi OS compatibility still needs to be checked on an actual Pi 500+. Browser access is also planned for ChatGPT.
Claude Code · Codex CLI · ChatGPT Linux · Hermes · Ollama · LM Studio
AUTHORS AND TERMS
A free image.
Visible component licences.
Our contribution is the integration, configuration and radio production workflow. Each tool’s authors retain their rights. This is a distribution plan for the main components, not a complete inventory of a finished image. Package versions, licences and source provenance will be recorded from the actual build.
| Component | Licence / terms | Delivery plan |
|---|---|---|
| Raspberry Pi OSOS and firmware | Per-package: GPL, LGPL, BSD, MIT and othersRaspberry Pi · Debian | Image base. Retain each package’s terms, required source and firmware notices. Full inventory after building. |
| Codex CLICoding agent | Apache-2.0LICENSE | Planned inclusion after ARM64 acceptance. The client licence does not include OpenAI services or model weights. |
| Claude CodeAgent with your own account | Proprietary / Anthropic termsLICENSE · Terms | User installation from the vendor. No assumed right to redistribute a bundled binary. |
| ChatGPTBrowser / Linux ARM64 preview app | Proprietary / OpenAI termsLinux app · Terms | Link to the service and optional vendor installation, using your own account. Desktop version needs a Pi test. |
| Hermes AgentAgent environment | MITNous Research / LICENSE | Installation candidate. Retain licence and attribution; models, integrations and their costs are separate. |
| OpenClawAgent and tool gateway | MIT + third-party noticesLICENSE · Pi guide | Installation candidate with user configuration. Individual skills and binaries need ARM64 acceptance. |
| Szpieg+ / L00P.AIArchive, transcription and playback | Free-edition terms pendingL00P.AI / Szpieg+ | Planned free image edition. Modules, user rights and the Pi port must be defined before release. No MIT assumption. |
| OllamaLocal model runtime | MITLICENSE · Linux | Linux ARM64, planned inclusion after CPU acceptance. Each downloaded model has separate terms. |
| llama.cppAlternative model runtime | MITLICENSE | ARM64 CPU candidate. Models and quantisations pinned with their own hashes and licences. |
| LM StudioOptional model interface | Proprietary / LM Studio App TermsApp Terms · System requirements | The app will not be bundled. Vendor link; redistribution requires separate permission. ARM64 needs testing on the Pi. |
| whisper.cpp / OpenAI WhisperASR runtime / model weights | MIT / MITwhisper.cpp / LICENSE · Whisper / LICENSE | Candidates: small, medium, large-v3-turbo. Retain runtime and weights licences separately; choose the export after testing. |
| sherpa-onnx / ONNX RuntimeAlternative ASR runtime | Apache-2.0 / MITsherpa-onnx / LICENSE · ONNX Runtime / LICENSE | ARM64 CPU candidate. Retain both licences and dependency notices; ASR weights have separate terms. |
| Parakeet TDT 0.6B v3NVIDIA ASR weights | CC-BY-4.0Model card · ONNX exports | Int8 candidate. Attribute NVIDIA, source, licence and export modifications. Benchmark the exported version separately. |
| Qwen3-ASR 0.6BQwen ASR weights | Apache-2.0Model card · ONNX exports | ONNX comparison candidate. Pin the export, provenance and modification notices before distribution. |
| Qwen3 1.7B / 4BSmall text models | Apache-2.01.7B · 4B | Q4 candidates for local testing. Weights, context and quantisation selected after measurement; no agent-speed promise. |
| FFmpegMedia processing | LGPL-2.1-or-later / GPL depending on buildLegal | Licence depends on build flags and included libraries. Inventory the actual binary; do not assign one licence to all codecs. |
Raspberry Pi OS contains packages under different licences. The release must retain attribution and licence texts and fulfil applicable source-code obligations. This also covers libraries, firmware, fonts and media tools. Debian / licences.
LM Studio permits personal and internal business use, but its current terms restrict redistribution of the app. We therefore plan an official vendor installation link. Claude Code and the ChatGPT app are also treated as separate vendor installations. Szpieg+’s stdlib approach reduces code dependencies; it is not itself a licence.
MEASUREMENT BEFORE A PROMISE
Local transcription.
Accuracy to be measured.
The Pi 500+ has 16 GB RAM, four 2.4 GHz Cortex-A76 cores and a 256 GB NVMe SSD as standard. That makes local ASR testing reasonable, but does not establish real-time speed or a particular error rate. The US variant denotes the keyboard, not a faster processor. Raspberry Pi 500+.
Whisper / whisper.cpp
Candidates: small, medium and large-v3-turbo. The runtime supports Raspberry Pi and quantisation. We will measure how much time improved accuracy costs on Polish recordings.
whisper.cpp →Parakeet TDT 0.6B v3
Candidate: an int8 export through sherpa-onnx. NVIDIA reports Polish WER of 7.31% on FLEURS and 7.28% on MLS. These are vendor results, not measurements on a Pi or our archive.
Model card →Qwen3-ASR 0.6B
Supports Polish; sherpa-onnx documents an ONNX export. It will be a third candidate for comparing accuracy and speed.
Model card →Initial text-model candidates: Qwen3 1.7B and 4B in Q4 form through llama.cpp / Ollama. We will start with short context and one task at a time. These are test candidates, not a performance guarantee or a claim of cloud-model quality. 1.7B · 4B.
How will we evaluate the below-4% WER target?
We will prepare at least two hours of human-verified reference material: studio, telephone, noise, overlapping speech, names and numbers. Test data will be kept separate from tuning data. We will record normalisation rules and raw results. Scribe will be another system to compare, not the ground truth.
WER = (S + D + I) / N — substitutions, deletions and insertions divided by reference words. We will also report name and number errors, results for each recording category, runtime, RAM and temperature. RTF below 1 means processing faster than the recording duration; it is a separate measure from WER.
We do not yet have a Pi 500+ measurement. Below 4% remains a target for a defined dataset, not a promise for any archive.
What must the first release include?
Base checked on 25 September 2026: Raspberry Pi OS Desktop 64-bit, released 15 September 2026, Debian 13 Trixie, kernel 6.18. Before building, we will check the stable release again and pin the exact file and checksum. Raspberry Pi OS.
- Verified SSD boot on a physical Pi 500+, Polish language and US keyboard; other layouts configurable.
- A complete inventory of actual versions, licences, sources and checksums; explicit terms for the free Szpieg+ edition.
- A clean first boot: none of our accounts, keys, recordings or client data; unique device identities.
- Archive selection, a transcription sample, playback from a word and export with a time map, tested on the Pi.
- Separate acceptance for offline and cloud paths, plus published benchmark results.
- Torrent, magnet link, a SHA-256 checksum on this page and instructions for writing the image to SSD. Test a restored copy of the downloaded release, not just the build device.
L00P.AI image price: free. Hardware, paid model services, support and custom deployments are separate. This is an independent L00P.AI project; listing tools does not imply a partnership with their vendors.