ComputeArena CLI 0.1.1

ComputeArena CLI 0.1.1

This release makes model identity verifiable. Reports now state exactly which
published model file was benchmarked, the client checks that claim against
Hugging Face before uploading, and computearena.ai verifies it again on its
own. It also adds update notices, server-driven upgrade guidance, and clearer
BaseRT support on Linux x86-64.

Install or upgrade

On macOS or Linux:

curl -LsSf https://computearena.ai/install.sh | sh

The same command upgrades an existing installation. Saved reports, the
installation signing key, login sessions, and installed runtimes are kept, and
the data directory gains a model-provenance/ folder for the receipts
described below. Release archives and checksums are also available below for
manual installation.

Highlights

• Verifiable model identity. Every new report carries a
  computearena-model/1 identity: the upstream Hugging Face model, the exact
  artifact (repository, immutable revision, file path, format, and
  runtime-specific quantization), the SHA-256 of the bytes that were
  benchmarked, and how they were obtained. The fields from 0.1.0 remain
  populated, and reports saved by earlier releases still submit.
• Managed downloads leave a receipt. GGUF downloads are pinned to an
  immutable revision, hashed while streaming, and checked against the Hub's
  published SHA-256. BaseRT models pulled through the picker are matched to
  their exact published file by hash. Models installed before this release are
  matched the same way the first time they are benchmarked; when no published
  file matches, the report records the repository and the hash rather than a
  guessed file name.
• BaseRT catalogue in the model picker. The BaseRT picker lists BaseRT's
  public catalogue and runs basert pull with a backend-aware target, so
  BaseRT keeps choosing the compatible artifact, downloading split files, and
  converting. Existing .base and .gguf files can still be entered directly.
• computearena identify. Bind a file you downloaded or copied yourself to
  one exact Hub file: computearena identify ./model.gguf <hf-file-url>
  succeeds only when the local SHA-256 matches the published object. A
  repository URL or a model name is not enough, and a later identify can
  correct an earlier receipt.
• Checks before submission. Before uploading, the client checks each
  report's artifact against Hugging Face and lists it as verified, unresolved,
  or check unavailable. A file whose hash contradicts the claimed published
  file is set aside as invalid; in the full-screen interface the remaining
  reports are still uploaded, and the summary says how many were left out.
  When computearena.ai has verified an artifact, its outcome is shown after
  the upload.
• Model families. A repository's base_model link counts only when the
  repository declares itself a quantization of that model. Finetunes,
  adapters, and merges stay their own model, so an instruct model is not folded
  into its pretraining base. computearena.ai decides the final grouping, and a
  difference between the client's record and that grouping never blocks a
  submission.
• Update notices and upgrade guidance. Interactive sessions check for the
  latest release in the background, keep the answer for a day, and mention a
  newer version when one exists; the check never blocks offline use. API
  requests carry the client version, and when the service retires a release
  the client prints the installed version, the minimum required, and the
  install command, then stops the submission batch.
• BaseRT on Linux x86-64. ComputeArena warns before entering BaseRT on
  Linux x86-64, where no prebuilt BaseRT runtime is published, and points to
  llama.cpp or a compatible harness you built yourself instead of an installer
  that cannot help there.

Supported systems

Prebuilt ComputeArena binaries are available for:

• macOS on Apple Silicon;
• Linux x86_64 with glibc 2.31 or newer; and
• Linux arm64 with glibc 2.31 or newer.

Windows and Intel Macs are not supported in this release. ComputeArena does not
bundle an inference runtime or model; install BaseRT or llama.cpp and obtain a
compatible model before benchmarking. Prebuilt BaseRT runtimes exist for macOS
on Apple Silicon and for Linux arm64 with CUDA; on Linux x86-64, use llama.cpp
or a BaseRT benchmark harness you built yourself.

The macOS binary is not signed with an Apple Developer ID. The release archive
is checksum-verified and carries a Sigstore keyless signature from the release
workflow. Browser-downloaded copies may need their quarantine attribute removed
as described in the README.

Model identity and trust

A verified artifact means the bytes that were benchmarked are a file published
on Hugging Face, and the model family follows what that file's publisher
declares. It does not prove that the publisher labelled the model correctly,
that the runtime executed those bytes honestly, or that two fine-tunes with
different lineage are equivalent. The client's checks are advisory;
computearena.ai repeats them independently, because a public client is not a
trust boundary. Files with no evidence, gated repositories the service cannot
read, and ambiguous merges stay unresolved rather than being grouped by file
name. Set HF_TOKEN in your environment to download from gated repositories.

Accuracy benchmarks, MLX, vLLM, Windows, and Intel macOS support are not part
of v0.1.1.

For usage and verification details, see the
README and
docs/model-identity.md.
Questions and feedback are welcome in the
ComputeArena Discord.

ComputeArena CLI 0.1.0

ComputeArena CLI 0.1.0

This is the first public release of the ComputeArena command-line client. It
runs versioned prefill and decode benchmarks through local inference
runtimes, keeps reports offline until you choose to submit them, and publishes
selected results to computearena.ai.

Install

On macOS or Linux:

curl -LsSf https://computearena.ai/install.sh | sh

The installer explains every change before proceeding, verifies the downloaded
archive, installs without sudo, and can configure ~/.local/bin on your
PATH. Release archives and checksums are also available below for manual
installation.

Highlights

• BaseRT and llama.cpp adapters. Benchmark .base models with
  basert-benchmark-harness or .gguf models with llama-bench. Runtime
  capability checks, executable discovery, model selection, and result
  translation stay isolated behind a common adapter interface.
• Interactive and scriptable workflows. Run computearena for the
  full-screen terminal interface, or use commands such as computearena basert run, computearena llama-cpp run, computearena list, computearena inspect, computearena verify, and computearena submit.
• Offline-first signed reports. Every report is saved locally and signed
  with an installation-specific Ed25519 key. Runtime and model files are hashed
  before and after execution, and a changed artifact prevents the report from
  being signed.
• Explicit submission control. Benchmarking does not require an account.
  When submitting, the client previews the exact public JSON, excludes local
  file paths, asks for confirmation, and handles valid reports separately from
  invalid ones.
• Automatic telemetry. ComputeArena records available process memory,
  temperature, power-state, NVIDIA, and ROCm observations while using one
  measurement pass with repeated samples rather than separate telemetry
  replays. A capability-gated BaseRT path is ready for native same-run energy,
  allocator, and KV-cache telemetry.
• Optional thermal conditioning. Standard runs start after warmup;
  thermally controlled runs add recorded cooldown waits. The plan shows the
  runtime-specific workload and time estimate before execution.
• Runtime and model setup. The client can find installed runtime binaries,
  guide BaseRT model pulls, search and download GGUF files from Hugging Face,
  and install supported prebuilt runtime bundles after showing the source,
  destination, and integrity check.
• Portable identity and storage. Conservative chip normalization keeps
  known aliases together without guessing unknown hardware. Existing staging
  data is migrated to the standalone ComputeArena data directory without
  changing saved reports, login sessions, signing keys, or installed runtimes.

Supported systems

Prebuilt ComputeArena binaries are available for:

• macOS on Apple Silicon;
• Linux x86_64 with glibc 2.31 or newer; and
• Linux arm64 with glibc 2.31 or newer.

Windows and Intel Macs are not supported in this release. ComputeArena does not
bundle an inference runtime or model; install BaseRT or llama.cpp and obtain a
compatible model before benchmarking.

The macOS binary is not signed with an Apple Developer ID. The release archive
is checksum-verified and carries a Sigstore keyless signature from the release
workflow. Browser-downloaded copies may need their quarantine attribute removed
as described in the README.

Benchmark and integrity notes

The default text benchmark records three repetitions for prefill sizes 128
through 16384 and 128 decode tokens. BaseRT and llama.cpp retain distinct
protocol identifiers, context semantics, runtime versions, backends, and
quantization metadata so unlike measurements are not silently presented as
equivalent.

A valid report signature proves that the report has not changed since the
installation signed it. It does not prove that the client, runtime, driver,
operating system, or machine reported truthful results. Executable and model
hashes identify the files that were checked around the run; they are not remote
attestation.

Accuracy benchmarks, MLX, vLLM, Windows, and Intel macOS support are not part of
v0.1.0.

For usage and verification details, see the
README. Questions and
feedback are welcome in the
ComputeArena Discord.