PRSN EX Verified
Signed, re-runnable checks on the AI models you run yourself.
A free service. We run narrow, re-runnable technical checks on specific open-weight AI model files you run locally on your Mac, and publish a signed record of exactly what we found.
We describe and report; we never recommend. We never host the files — every record links to the exact file at its original source, hash-pinned, so you can confirm byte-for-byte that what you downloaded is what we tested.
What the model says is its own character; we test the container, not the conversation.
PRSN EX Verified.
Verified: file integrity · format safety · offline operation — not behavior, content, or safety.
Verified model files
Every record is backed by a cryptographically signed manifest — verify any of it yourself.
All performance figures are measured, not estimated, on one machine: a Mac Mini with the Apple M4 Pro chip and 64 GB of memory.
Results on other Apple Silicon Macs will differ.
Reviewed as of the date shown; we don't promise a fixed update interval.
| Model | Mac RAM needed | Tokens/sec | Reviewed | Report | Source |
|---|---|---|---|---|---|
| Gemma-4-E4B-it (QAT UD) GGUF · text + vision Model by Google · Converted by unsloth Apache 2.0 · free for commercial use | 8 GB (to 32k context)16 GB (full 128k context) | ~65 tok/s | 2026-07-16 | View at source | |
| Gemma-4-12B-it (QAT) GGUF · text + vision Model by Google · Converted by lmstudio-community Apache 2.0 · free for commercial use | 16 GB | ~30 tok/s | 2026-07-15 | View at source | |
| Gemma-4-12B-it (QAT) MLX† · text + vision Model by Google · Converted by mlx-community Apache 2.0 · free for commercial use | 24 GB (measured to 32k context depth) | ~30 tok/s | 2026-07-16 | View at source | |
| Gemma-4-26B-A4B-it (QAT UD) GGUF · text + vision Model by Google · Converted by unsloth Apache 2.0 · free for commercial use | 24 GB (everyday)32 GB (full 256k context) | ~74 tok/s | 2026-07-16 | View at source | |
| Qwen3.6-27B GGUF · text + vision Model by Alibaba · Converted by unsloth Apache 2.0 · free for commercial use | 32 GB (everyday)48 GB (256k context) | ~11 tok/s | 2026-07-15 | View at source | |
| Qwen3.6-35B-A3B (UD) GGUF · text + vision Model by Alibaba · Converted by unsloth Apache 2.0 · free for commercial use | 48 GB (all context depths, with headroom) | ~46 tok/s | 2026-07-16 | View at source | |
| Qwen3.6-35B-A3B (UD) MLX† · text + vision Model by Alibaba · Converted by unsloth Apache 2.0 · free for commercial use | 48 GB (measured to 32k context depth) | ~74 tok/s | 2026-07-16 | View at source | |
| Qwen3.8-27B GGUF · text + vision Model by Alibaba · Converted by unsloth Apache 2.0 · free for commercial use | 32 GB (everyday, 32k context)48 GB (full 256k context depth) | ~10 tok/s | 2026-08-16 | View at source | |
| Muse-Glimmer-30B GGUF · text + vision Model by Meta · Converted by lmstudio-community Apache 2.0 · free for commercial use | 32 GB (all depths, to the full 128k context) | ~12 tok/s | 2026-08-16 | View at source |
- Mac RAM needed
- 8 GB (to 32k context)16 GB (full 128k context)
- Tokens/sec
- ~65 tok/s
- Reviewed
- 2026-07-16
- Mac RAM needed
- 16 GB
- Tokens/sec
- ~30 tok/s
- Reviewed
- 2026-07-15
- Mac RAM needed
- 24 GB (measured to 32k context depth)
- Tokens/sec
- ~30 tok/s
- Reviewed
- 2026-07-16
- Mac RAM needed
- 24 GB (everyday)32 GB (full 256k context)
- Tokens/sec
- ~74 tok/s
- Reviewed
- 2026-07-16
- Mac RAM needed
- 32 GB (everyday)48 GB (256k context)
- Tokens/sec
- ~11 tok/s
- Reviewed
- 2026-07-15
- Mac RAM needed
- 48 GB (all context depths, with headroom)
- Tokens/sec
- ~46 tok/s
- Reviewed
- 2026-07-16
- Mac RAM needed
- 48 GB (measured to 32k context depth)
- Tokens/sec
- ~74 tok/s
- Reviewed
- 2026-07-16
- Mac RAM needed
- 32 GB (everyday, 32k context)48 GB (full 256k context depth)
- Tokens/sec
- ~10 tok/s
- Reviewed
- 2026-08-16
- Mac RAM needed
- 32 GB (all depths, to the full 128k context)
- Tokens/sec
- ~12 tok/s
- Reviewed
- 2026-08-16
These models reason before they answer. Run one with its chat template and enough room to respond — with too tight an output limit, the reply can come back empty. Each report gives the specifics.
† MLX files were tested with one specific version of the software that runs them; the report names it.
Verify it yourself
You never have to take our word for it. Each report publishes the file's full SHA-256 and the exact commands to check it. On your machine:
# compare your downloaded file against its record
shasum -a 256 your-model-file.gguf
# the result must match the SHA-256 printed in that model's reportA one-command helper that runs the full check — the manifest's signature first, then the signed withdrawal list, then the bytes, refusing on any mismatch — ships with our public repository. Helper script license & notice. Repository: coming soon.
What we do not test
What we do not test: how the model behaves. Whether its answers are good, safe, fair, or filtered is the model's own character, and our records make no claim about it.
The fine print
The documents that govern this service: