PPLX Embed v2

Native Swift/MLX inference loads the original Perplexity FP32 safetensors directly. No Python inference process, remote code execution, or conversion is required.

Managed model	Output	Input
text-embed-pplx-v2-late-0.6b	Normalized 128-dimensional token vectors	Text queries, text or image documents
text-embed-pplx-v2-late-9b	Normalized 128-dimensional token vectors	Text queries, text or image documents
text-embed-pplx-v2-context-9b-preview	Int8-valued 2048-dimensional chunk vectors	Text queries and document chunks
text-embed-pplx-v2-late-9b-mixed-4bit	Same late token-vector contract	Text queries, text or image documents
text-embed-pplx-v2-context-9b-preview-8bit	Same contextual chunk-vector contract	Text queries and document chunks

The two Sawfwair 9B variants contain ready-to-load packed MLX weights. Users pull them directly; no conversion step is needed. Late uses Q4 transformer projections and Q8 token embeddings. Context uses Q8 for both because the Q4 candidate showed more vector drift. Gates, convolution, norms, vision, and final projections remain FP32. Contextual int8 output is independent of weight packing.

mere.run model pull text-embed-pplx-v2-late-9b-mixed-4bit
mere.run text embed "What drives photosynthesis?" \
  --model text-embed-pplx-v2-late-9b-mixed-4bit --task query
mere.run model pull text-embed-pplx-v2-context-9b-preview-8bit
mere.run text embed --chunks-json chunks.json \
  --model text-embed-pplx-v2-context-9b-preview-8bit --dimensions 1024 --normalize

Admission checks available memory before inference; active workloads can temporarily block a run. The checkpoint's maximum context is not a fit guarantee. See the qualification record for paired output drift and the bounded validation scope.

Install explicitly before inference; these models do not download automatically:

mere.run model pull text-embed-pplx-v2-late-0.6b
mere.run text embed "what statute governs limitations?" \
  --model text-embed-pplx-v2-late-0.6b --task query --output query.json
mere.run text embed "A document passage" \
  --model text-embed-pplx-v2-late-0.6b --task document --output document.json
mere.run text embed --image ./document.png \
  --model text-embed-pplx-v2-late-0.6b --task document --output image.json

Late models keep a vector for every retained token, including the initial query or document marker. Document punctuation token IDs are excluded using the checkpoint's mask configuration; queries retain them. Query/document limits are 1024/4096 tokens. Text truncates from the right; --max-tokens can lower the limit. Image batches must run separately from text batches and must use the document task. Images use the checkpoint's Qwen3-VL resize settings, RGB bicubic resampling, vision tower, and multimodal rotary positions. An image exceeding the token limit is rejected intact; resize it before encoding.

Use sum-MaxSim to compare late vectors: for each query token, find its largest dot product with any document token, then sum those maxima. The Swift API exposes PPLXEmbedV2Result.maxSim(query:document:). The two late models share the upstream embedding space, allowing small-model queries against a large-model index.

For contextual encoding, save chunks.json:

{"documents":[["Curiosity begins in childhood.","Scientific breakthroughs often start with a question."],["The curiosity rover explores Mars."]]}

mere.run model pull text-embed-pplx-v2-context-9b-preview
mere.run text embed --chunks-json chunks.json \
  --model text-embed-pplx-v2-context-9b-preview --task document \
  --dimensions 1024 --normalize --output chunks-embeddings.json
mere.run text embed "What drives scientific breakthroughs?" \
  --model text-embed-pplx-v2-context-9b-preview --task query --normalize

Chunks in each document share a forward pass. Pooling uses each chunk's overlapping BPE token span, excluding prefix/separator-only tokens. NFC-normalized bytes preserve chunk boundaries for decomposed Unicode. Queries mean-pool all valid tokens, including the query marker. Empty chunks return zero vectors. Document prefixes tokenize as ordinary BPE; non-special added tokens such as the real context checkpoint's chunk separator remain intact, matching upstream split_special_tokens=True. Query markers remain explicit special IDs.

Context vectors use FP32 projection, round(tanh(x) * 127), and int8 clamping. They are unnormalized by default and represented as integer-valued JSON numbers. Compare them with cosine similarity, or request --normalize and use dot products. The trained dimensions are 1024 and 2048; truncation occurs before normalization. The 262144-token limit rejects oversized inputs without truncation or automatic chunk splitting. Available memory can impose a lower practical limit.

--task defaults to document for PPLX. These results have a separate multi-vector JSON schema: representation, dimensions, normalized, and data:[{index,embeddings:[[...]],tokenCount}]. They are CLI/Swift API models; the OpenAI single-vector /v1/embeddings route does not serve them.

The contextual model is a preview. Keep its embeddings separate from future checkpoint versions.

Pins and validation

The original checkpoints pin immutable upstream revisions:

• Late 0.6B: dd4e95b836a73f6f0c32e46ea127c0b86b02169e
• Late 9B: 77e936a1b18ed2ac00b7c76fccd70dc6a1bb1c18
• Context preview: b667039ee8b438a6350fbc91bbcecd86f9d363ba

The packed managed entries pin late mixed Q4/Q8 and context Q8. Full-checkpoint paired FP32 diagnostics and native CLI runs are recorded in the qualification report.

Synthetic FP32 fixtures exported from Transformers 5.4.0 cover the hybrid encoder, bidirectional full attention, and vision path. Tests also cover marker/mask semantics, chunk spans, projection/quantization, dimensions, input rejection, model routing, and CLI parsing. Broad retrieval quality, cross-model alignment, and large-context memory behavior require separate qualification. The bounded full-checkpoint diagnostics and synthetic fixtures do not establish those properties.

Upstream: collection, late model, and context preview.