Appearance
PPLX Embed v2 9B qualification — October 9, 2026
Both packed checkpoints passed bounded native public CLI checks. No Python inference process is used locally. Users pull the packed artifacts directly; conversion is a publisher task.
| Checkpoint | Weight files | FP32 vs packed mean cosine | Minimum cosine | Retrieval top choices |
|---|---|---|---|---|
| text-embed-pplx-v2-late-9b-mixed-4bit | 7.32 GB | 0.999065 | 0.991924 | 8/8 agree |
| text-embed-pplx-v2-context-9b-preview-8bit | 10.81 GB | 0.999873 | 0.999813 | 8/8 agree |
These short-input checks do not establish maximum context or concurrent residency. Original FP32 9B managed entries retain their separate admission guidance.
Late uses affine Q4/group-64 transformer projections and Q8/group-64 token embeddings. Context uses Q8/group-64 for both. Recurrent gates, convolution, norms, vision, and final projections retain FP32. Context's int8-valued output contract is independent of weight quantization.
The context Q4 candidate kept the same eight top choices but had mean cosine 0.980567 and minimum cosine 0.970761. Q8 improved these to 0.999873 and 0.999813, so Q8 was selected. Selection used this same small suite: do not interpret it as an independent heldout benchmark. quality_qualified remains false.
What was measured
Conversion used MLX CUDA 0.32.2. Independent PyTorch integer unpacking was checked against MLX dequantization samples, then FP32 and reconstructed packed weights ran with Torch 2.8.0 and Transformers 5.4.0. The multilingual suite includes eight queries and eight documents, marker/Unicode/punctuation cases, contextual empty chunks and shared chunk boundaries, and one synthetic image for late. There are 19 late and 20 context cases. Both profiles pass minimum cosine 0.95, mean cosine 0.98, finite vectors, and complete top-choice agreement gates. The legacy late report field passes_heldout_gates names these diagnostic gates; it does not establish broader heldout quality.
Native CLI validation compares token/vector counts and every output vector against the exported reconstructed-weight reference. Late covers text and a synthetic image; context covers query/document chunks. A separate real-tokenizer fixture verifies NFC-normalized decomposed accents, Hangul, empty chunks, emoji, and literal markers against Transformers token IDs and offsets. A further full-checkpoint check passed 1024-dimensional normalization across all 11 document cases, preserving the empty chunk as a zero vector.
Broad retrieval benchmarks, image retrieval quality, small/large cross-model alignment, maximum-context behavior, and future context-preview compatibility remain unqualified. Context-preview embeddings must stay separate from later checkpoint versions.
Artifacts and reproduction
- Sawfwair/pplx-embed-v2-late-9b-MLX-Mixed-4bit, source
perplexity-ai/pplx-embed-v2-late-9b@77e936a1b18ed2ac00b7c76fccd70dc6a1bb1c18. CUDA diagnostic, native proof, checksums. - Sawfwair/pplx-embed-v2-context-9b-preview-MLX-8bit, source
perplexity-ai/pplx-embed-v2-context-9b-preview@b667039ee8b438a6350fbc91bbcecd86f9d363ba. CUDA diagnostic, native proof, checksums.
See the conversion runbook and native runtime guide. Each artifact preserves its actual converter, source provenance, upstream license information and model card, and weight reconstruction measurements. Publication staged privately, verified every file's size and SHA-256, then verified the public immutable revision anonymously. The Hugging Face write token stayed on the Mac. The conversion pod was terminated after all selected weights and evidence were preserved locally; its attached volume was removed with it.
The JSON record contains immutable pins, per-batch parity, measured verified file hashes, and the labeled allocation-cost estimate. Machine metadata is omitted from this derived record; the linked immutable artifacts preserve the original evidence.
Repository validation
The full repository gate (./scripts/check.sh) passed on the isolated PPLX branch against main: 5370 XCTest cases (442 skipped, zero failures), 192 Swift Testing cases, build, lint, policy checks, CLI help checks, and hygiene scans. The documentation build also passed. Real-tokenizer fixtures separately passed for both 9B tokenizers.
Both managed model pull commands recognize their complete local installs. The full-checkpoint native runs above passed with normal admission. A later managed-ID inference retry was blocked by available-memory admission. Other active workloads can affect admission even when an individual checkpoint run has passed.