tiyuvta inference
Menu
embeddings · live · batch-class

Qwen3-Embedding-8B

Specifically designed for text embedding and ranking tasks. Building upon the dense foundational models of the Qwen3 series... This series inherits the exceptional multilingual capabilities, long-text understanding, and reasoning skills of its foundational model.

Qwen’s own description of the model, from Qwen3-Embedding-8B model card. Served here as qwen/qwen3-embedding-8b.

What Qwen built it for

in the vendor’s words

Quoted from Qwen3-Embedding-8B model card, retrieved 2026-08-27.

Published benchmarks

Qwen’s figures, not ours
Qwen3-Embedding-8B — scores as published by Qwen, retrieved 2026-08-27
Benchmark What it measures Score How Qwen states it
MTEB multilingual leaderboard Text embedding, multilingual 70.58 The card states the rank and score as of June 5, 2025: "ranks No.1 in the MTEB multilingual leaderboard".

Every score above is Qwen’s own published result for this model, copied from Qwen3-Embedding-8B model card on 2026-08-27 — we did not run these benchmarks and we do not restate them as our own. Each vendor ran its own harness with its own judge, so these tables are not a league table: scores from two different cards are not comparable, even where the benchmark name matches.

Recommended usage

what Qwen recommends

Instructions

“using instructions (instruct) typically yields an improvement of 1% to 5% compared to not using them. Therefore, we recommend that developers create tailored instructions specific to their tasks and scenarios.”

“In multilingual contexts, we also advise users to write their instructions in English, as most instructions utilized during the model training process were originally written in English.”

On this endpoint: prepend the vendor’s instruction format to QUERY texts yourself (Instruct: <task>\nQuery: <text>); documents are embedded as-is.

Dimensions

Native vectors are 4096-dimension, L2-normalized. The model is trained matryoshka-style: the request’s dimensions field truncates to any size from 32 to 4096, re-normalized — smaller indexes, same API call.

Context length

Qwen gives the window as 33K tokens per input, which is what we serve — long documents go in whole instead of pre-truncated.

What to point it at

Qwen names these capabilities for the series:

  • “text retrieval, code retrieval, text classification, text clustering, and bitext mining”
  • “support for over 100 languages”, including programming languages

Recommendations on this page are Qwen’s, from Qwen3-Embedding-8B model card, retrieved 2026-08-27.

What it does on this endpoint

our gates · not vendor figures
exactness cosine ≥ 0.998 vs the fp16 reference, on the deployed artifact
vectors 4096-dim L2-normalized, MRL-truncatable to 32
context 33K tokens per input

No speed figure is published for this model on purpose: throughput is best-effort by contract (see the price section). What IS gated, on every deploy, is byte-level correctness — the served vectors are checked against the fp16 transformers reference for cosine and retrieval-rank parity on the exact artifact and binary serving, before cutover.

Price, and the scheduling contract behind it

per million tokens · input only · no request fee
Token rates — Qwen3-Embedding-8B
Input Output
$0.01$0.00 — an embedding request generates nothing

This is a batch-class surface by design: embedding traffic rides the serving valleys of the generation models and is shed with a retryable 429 when their lanes are busy. The paying interactive lanes are never taxed — and that subordinate scheduling is exactly what this price buys. Retry with backoff; throughput is best-effort, correctness is gated.

One price, no tiers, prepaid credit — the same balance as every other model here. Full price sheet.

Capabilities

qwen/qwen3-embedding-8b

OpenAI wire format

The standard /v1/embeddings schema: a string or array of strings in, vectors out. Point an existing OpenAI client at https://api.tiyuvta.ai/v1 and set the model id.

Same key, same balance

One base URL, https://api.tiyuvta.ai/v1; the model field routes. A RAG stack runs embedding, retrieval and generation on a single prepaid credit — no second vendor, no second bill.

Deterministic output

The same (input, dimensions) pair returns the same vector, byte-for-byte — safe to cache on your side, and index migrations diff cleanly.

Pairs with

Qwen3-Reranker-8B — embed with this, rerank with it: the complete retrieval stack from one vendor family, on one key.

First request

qwen/qwen3-embedding-8b
shell — embed https://api.tiyuvta.ai/v1 
curl https://api.tiyuvta.ai/v1/embeddings \
  -H "Authorization: Bearer $TIYUVTA_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen/qwen3-embedding-8b",
       "input": ["first passage", "second passage"],
       "dimensions": 1024}'

dimensions is optional — omit it for the full 4096.

shell — the retry loop batch-class 
# a 429 means the interactive lanes are busy — retry with backoff
for i in 1 2 3 4 5; do
  code=$(curl -s -o out.json -w "%{http_code}" https://api.tiyuvta.ai/v1/embeddings \
    -H "Authorization: Bearer $TIYUVTA_KEY" \
    -H "Content-Type: application/json" -d @request.json)
  [ "$code" != 429 ] && break
  sleep $((i * i))
done

A 429 is the batch-class contract working, not an outage: the interactive lanes were busy and your request should come back with backoff.