Intel Arc AI Stack Compatibility Tracker

Evidence-first compatibility checks for Intel Arc B580 AI stacks. arcor helps developers, platform engineers, and AI agents decide what is known, what is supplied by the caller, and what still needs verification. It combines explicit stack configurations, live Intel/GitHub release metadata, model-catalog lookups, and user-selected GitHub issues into a consistent, machine-readable dataset.

The actor is deliberately cautious. It never invents a fallback release, never turns missing evidence into a positive compatibility verdict, and never presents a stale cached response as a fresh source check. A result can be live, cached, input, or error; the source URL, source timestamp, run timestamp, confidence, and evidence are included whenever they are available.

When should an AI agent use this?

Use this actor when a question involves one of the following:

• “Can this Ollama, llama.cpp, ComfyUI, or Intel software configuration run on an Arc B580?”
• “What is the latest relevant Intel release metadata, and does the release mention Arc hardware?”
• “Is this GitHub issue still open, and has its status changed since the previous run?”
• “Does this model identifier exist in the Hugging Face catalog, and what compatibility evidence is actually available?”
• “Give me a structured, source-aware result for an automated deployment or monitoring pipeline.”

The actor is not an official Intel product and does not replace a hardware test. It is a decision-support and evidence-collection tool. A configuration marked unknown means exactly that: the available evidence is insufficient for a reliable yes/no verdict.

Keywords

Intel Arc, Arc B580, B580, AI compatibility, Ollama, llama.cpp, ComfyUI, IPEX, Intel oneAPI, XPU, GGUF, model compatibility, GitHub issue monitor, release monitor, evidence, AI agent, Docker, Ubuntu, local LLM, inference, structured output.

Quick start

The default demo input is deliberately local and does not call external APIs:

{
  "monitorMode": "dataset_only",
  "seedConfigs": [
    {
      "gpu": "Arc B580",
      "os": "Ubuntu 24.04 LTS",
      "driver": "Linux 6.8+",
      "aiEngine": "ollama",
      "engineVersion": "0.5.7",
      "modelFormat": "GGUF INT8",
      "works": "unknown"
    }
  ]
}

The result contains a stable configuration ID, a source status, the normalized stack fields, and an assessment object. Change works to confirmed, partial, or failed only when that status is supported by your own test evidence. Add source and verifiedAt to make the evidence auditable.

For a live release check:

{
  "monitorMode": "live",
  "modelQuery": "qwen2.5"
}

For a tracked issue, use a real GitHub issue reference:

{
  "monitorMode": "bugs_only",
  "trackBugIds": ["owner/repository#123"],
  "includeUnchangedBugs": false
}

The actor accepts owner/repository#number and full github.com/.../issues/... URLs. Pull requests and unsupported Intel-specific identifiers are returned as explicit source errors rather than being mislabeled as bugs.

Monitor modes

• live fetches current intel/llm-scaler release metadata and the latest tag from Intel’s archived intel-extension-for-pytorch repository. It performs an optional model lookup when modelQuery is set.
• dataset_only assesses the configurations supplied in seedConfigs and optionally looks up a model. It does not inject demo configurations.
• bugs_only fetches only the explicitly supplied GitHub issues and compares their state with the previous successful run.
• full combines live releases, supplied configurations, tracked issues, and an optional model lookup. Each source is deduplicated by its canonical ID.

includeUnchangedBugs is true by default so a current snapshot is returned. Set it to false for change-oriented monitoring; a new or updated issue remains visible, and a transition from open to closed is recorded as a resolved signal.

Input fields

Field	Type	Default	Purpose
monitorMode	enum	full	Selects the enabled source groups.
seedConfigs	array	[]	Up to 50 explicit GPU/OS/driver/engine configurations.
trackBugIds	array	[]	Up to 10 GitHub issue references.
arcSeries	enum	Arc B580	Selects the verified hardware scope. Arc B580 12GB is normalized to the same scope.
aiEnginesFilter	enum array	all four engines	Restricts configuration and model output to selected engines.
includeAssessment	boolean	true	Adds a cautious verdict and evidence list.
includeUnchangedBugs	boolean	true	Includes unchanged tracked issues in the output.
modelQuery	string	empty	Matches a model name or alias such as minimax h3 video, llama3.2, or qwen2.5.

Each seedConfigs item requires gpu, os, driver, aiEngine, engineVersion, and modelFormat. Optional fields include flags, customNodeConflicts, envVars, source, and verifiedAt. The supported evidence states are confirmed, partial, failed, and unknown; omission is equivalent to unknown, not confirmed. A valid source reference and parseable verification date are required before a result receives high confidence; invalid provenance is downgraded.

Output records

Every default-dataset item has these common fields:

• type: release, config, bug, model, or source_error.
• recordId: a stable hash-based ID for deduplication across runs.
• runId, runDate, and mode: run metadata.
• source, sourceUrl, sourceStatus, sourceUpdatedAt, and lastChecked: provenance and freshness.
• data: the record-specific payload.

Configuration assessments use canRun values yes, partial, no, or unknown, plus an English reason, estimated performance, confidence, and evidence list. An omitted works value never produces yes. EXL2 and AWQ are not treated as verified formats merely because another field says the stack works.

Release records distinguish what a source explicitly mentions from what remains unverified. A release that mentions Arc Pro B70 does not automatically claim Arc B580 support. The Intel Extension for PyTorch tag is included as a release signal, but its archived repository does not provide the Arc B580 evidence required for a positive compatibility verdict.

Bug records contain the real GitHub title, current status, labels, timestamps, relevance markers, and changeType (new, updated, or unchanged). The actor does not output the issue body, so usernames, local paths, tokens, and unrelated personal details are not copied into the result. Invalid references, unavailable sources, rate limits, and pull requests become source_error records with a code and retryability flag.

Source and failure behavior

Live requests use bounded timeouts, retry transient HTTP failures, validate response shapes, and cache sanitized source responses in a named key-value store. Release, issue, and model records expose data.sourceFetchedAt and data.sourceError, so a cached response is explicitly marked cached and can be distinguished from a fresh check. A source failure is never replaced with a fabricated release or issue.

GitHub’s public API can be rate-limited, especially when many users share an egress address. User-supplied issue lookups always use unauthenticated public requests, so an Actor-level GITHUB_TOKEN is never sent to a caller-selected repository. For higher-volume fixed-source lookups, configure an optional GITHUB_TOKEN in the Actor environment. The actor does not require a token for the small public lookups used by the default tasks. Review the source terms and rate limits before scheduling high-frequency runs.

Example tasks

1. Assess an Ollama Arc stack — use dataset_only with one Arc B580 configuration and includeAssessment: true.
2. Check current Intel AI release metadata — use live with an empty modelQuery.
3. Look up a model identifier — use live or dataset_only with modelQuery: "minimax h3 video".
4. Monitor an issue — use bugs_only with one or more explicit GitHub issue references and includeUnchangedBugs: false.

Limitations and responsible use

The verified hardware scope in this release is Arc B580 on the supplied Ubuntu stack. Other Arc series, Windows stacks, future drivers, and unreleased model combinations are not silently generalized. Release metadata can change after a run, and GitHub issue relevance is marked as evidence-based rather than guaranteed. Always test a deployment on the exact driver, model file, quantization, and runtime you intend to operate.

The actor reads public GitHub and Hugging Face endpoints. It does not scrape private repositories, bypass authentication, or make financial, legal, medical, or safety determinations. User-supplied environment variables and evidence fields are returned as part of the caller’s configuration; do not put secrets in them.

Related Actors

• Apify Open Source jobs scraper for developer-job data.
• Open Source Job Search for broader job-search workflows.
• GitHub Issues for general-purpose issue collection without Intel-specific assessment.

Changelog

0.2.0

• Replaced fabricated demo and fallback data with real source validation and explicit error records.
• Added canonical GitHub issue parsing, pull-request rejection, cross-run state, and resolved signals.
• Added safe tri-state evidence handling, engine/GPU filters, normalized model aliases, and live model existence checks.
• Added stable IDs, deduplication, bounded retries, normalized source caching, and per-item push/charge isolation.
• Aligned the Apify input schema, Zod parser, output schemas, README, and Store metadata.
• Added regression tests for URLs, source failures, assessment safety, model matching, and schema contracts.