Inconsistent Embedding Outputs with qwen3-embedding-0.6B

Problem Description

When vectorizing the same batch of inputs using qwen3-embedding-0.6B, the model occasionally returns two different embeddings:

• Majority of runs: Result A
• Minority of runs (~5%): Result B

Although the cosine similarity between A and B is very close to 1, the fact that the first embedding vector sometimes changes is both curious and concerning for downstream stability.

Environment

• OS: Centos7
• Model: qwen3-embedding-0.6B
• TEI Version: 1.7.2
• Client: Postman (HTTP POST)
• Hardware: A100 80GB
• Observed Frequency: ~5% of invocations return variant B

Reproduction Steps

1. Prepare the JSON payload:

   {
     "inputs": [
       "\n数据建模规范：设计规范/规范/标签",
       "频率（i_style） 业表 rdm_[业务写]_[标名] [业名][标名] 运好（or）"
     ]
   }

2. Send the request to the embedding endpoint in Postman.
3. Repeat the request 50–100 times and collect all returned embeddings.

Expected Behavior

• Consistency: The same input should always produce the identical embedding vector, ensuring reliable downstream retrieval and clustering.

Actual Behavior

• Two distinct outputs:
  • Result A (≈95% of runs)
  • Result B (≈5% of runs)

Outputs

Only the first 5 dimensions are shown for brevity; actual embeddings are 1024 dimensions long.

Example A (Majority)

[
  [
    -0.033525728,
    0.002407345,
    -0.008492636,
    -0.014752968,
    0.035222610,
    …
  ],
  [
    -0.033656940,
    -0.036949120,
    -0.006955928,
    -0.057693046,
    0.023221051,
    …
  ]
]

Example B (Minority)

[
  [
    -0.033118628,
    0.002002977,
    -0.008502096,
    -0.014664467,
    0.035161767,
    …
  ],
  [
    -0.033656940,
    -0.036949120,
    -0.006955928,
    -0.057693046,
    0.023221051,
    …
  ]
]