nomic-embed-text-v1.5, no MMTEB score in the table used
Vast.ai A10 on-demand median
Services
Embedding 1M documents of 500 tokens once a month (500M tokens), Qwen3-Embedding-8B (MMTEB 70.58) on a Vast.ai A10 / A10G costs about $7.84 in GPU rental, against $65 for OpenAI text-embedding-3-large and $10 for 3-small. The GPU price is a marketplace median. The speed is an estimate.
Enter your corpus size below. The calculator picks an open embedding model and a GPU, then compares that bill with OpenAI, Voyage, Cohere, Google, and hosted open models.
$7.84
this month · $0.0157 per 1M tokens · or $179.09 if the GPU stays on
MMTEB 70.58MTEB English v2 75.22
The pick is the highest MMTEB score among setups whose bill is no more than twice the cheapest scored setup, or $10 more, whichever is larger. Models without a score are priced but not picked.
Quality is the MTEB Multilingual mean over tasks (MMTEB, higher is better), as listed in the Qwen3-Embedding model card's copy of the MTEB leaderboard. English v2 is shown where the same table has it.
nomic-embed-text-v1.5, no MMTEB score in the table used
Vast.ai A10 on-demand median
Free, from this page
One email with this page’s inputs, the recommended model, GPU, and provider, a deployment checklist, and whether self-hosting or the API is the smaller bill.
Embedding deployment plan
Qwen3-Embedding-8BVast.ai · NVIDIA A10 / A10G
$7.84 this month
OpenAI 3-large $65.00 for the same work. GPU rental only.
500M tokens a month on OpenAI text-embedding-3-small is $10, and on text-embedding-3-large $65. As a batch job, Qwen3-Embedding-8B on a Vast.ai A10 / A10G is $7.84 of GPU time. Hosted open models sit in between: DeepInfra serves BGE-M3 for $5.
Kept on for all 730 hours to answer queries, the same GPU is $179.09 a month, more than text-embedding-3-large for this corpus. Self-hosting embeddings pays when the GPU is busy: large or frequently refreshed corpora, or text that cannot leave your network. Vector storage is not in either figure.
Quality is the MTEB Multilingual mean over tasks (MMTEB, higher is better), as listed in the Qwen3-Embedding model card's copy of the MTEB leaderboard. English v2 is shown where the same table has it. Scores link to the MTEB leaderboard. A dash means the model has no row in that table.
| Model | MMTEB | MTEB Eng v2 | Params | GPU | Provider | $/hour | Always-on / month | This workload | Per 1M tokens | Tokens/s per GPU |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-Embedding-8B · recommended Apache-2.0 | 70.58 | 75.22 | 7.57B | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $7.84 | $0.0157 | 4.3k estimate |
| Qwen3-Embedding-4B Apache-2.0 | 69.45 | 74.6 | 4.02B | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $4.16 | $0.0083 | 8.2k estimate |
| Qwen3-Embedding-0.6B Apache-2.0 | 64.33 | 70.7 | 596M | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $1.55 | $0.0031 | 22k estimate |
| multilingual-e5-large-instruct MIT | 63.22 | 65.53 | 560M | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $1.40 | $0.0028 | 24k estimate |
| gte-Qwen2-7B-instruct Apache-2.0 | 62.51 | 70.72 | 7.61B | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $7.88 | $0.0158 | 4.3k estimate |
| BGE-M3 MIT | 59.56 | — | 568M | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $1.42 | $0.0028 | 24k estimate |
| E5-Mistral-7B-instruct MIT | — | — | 7.11B | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $7.37 | $0.0147 | 4.6k estimate |
| jina-embeddings-v3 CC-BY-NC-4.0 · Non-commercial licence. Commercial use needs a licence from Jina. | — | — | 572M | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $1.43 | $0.0029 | 24k estimate |
| nomic-embed-text-v1.5 Apache-2.0 | — | — | 137M | NVIDIA A10 / A10G | Vast.ai A10 on-demand median | $0.2453 median | $179.09 | $0.34 | $0.0007 | 100k estimate |
“Measured” is the vLLM benchmark of E5-Mistral-7B. “Approx.” is a figure Spheron publishes as approximate. Everything else is a TFLOPS and parameter-count estimate.
List prices checked November 1, 2026. Batch discounts and free tiers are not subtracted unless the row says so.
| API | Type | $ per 1M tokens | This month | MMTEB | Note |
|---|---|---|---|---|---|
| OpenAI text-embedding-3-small | Managed API | $0.02 | $10.00 | — | |
| OpenAI text-embedding-3-large | Managed API | $0.13 | $65.00 | 58.93 | |
| Voyage voyage-4-large | Managed API | $0.12 | $60.00 | — | The first 200M tokens are free and the batch API is 33% off. Neither is subtracted. |
| Voyage voyage-4 | Managed API | $0.06 | $30.00 | — | |
| Voyage voyage-4-lite | Managed API | $0.02 | $10.00 | — | |
| Google gemini-embedding-001 | Managed API | $0.15 ($0.075 batch) | $75.00 | 68.37 | Score row is gemini-embedding-exp-03-07, the experimental release before gemini-embedding-001. |
| Cohere Embed 4 | Managed API | $0.12 | $60.00 | — | cohere.com/pricing no longer lists a per-token Embed price. $0.12 per 1M is the AWS Marketplace listing and matches LiteLLM's price table, read 1 November 2026. The 61.12 MMTEB row is Cohere multilingual v3, a different model, so it is not shown here. |
| DeepInfra BGE-M3 | Hosted open model | $0.01 | $5.00 | 59.56 | Score of the open BGE-M3 weights. |
| DeepInfra multilingual-e5-large-instruct | Hosted open model | $0.01 | $5.00 | 63.22 | Score of the open multilingual-e5-large-instruct weights. |
| Fireworks Qwen3-Embedding-8B | Hosted open model | $0.1 | $50.00 | 70.58 | Score of the open Qwen3-Embedding-8B weights. |
On-demand hourly price and the always-on month (hourly × 730). Community and non-preemptible rates are shown only where that provider publishes them. Each price links to the provider’s pricing page, checked November 1, 2026. A dash means that provider does not list this GPU.
| GPU | RunPod | Lambda | Modal | Vast.ai | AWS | Google Cloud |
|---|---|---|---|---|---|---|
| L4 | — | Nvidia L4 $0.7992/hr Preemptible $583.42/mo Non-preemptible $2.3976/hr · $1,750/mo checked November 1, 2026 | ||||
| A10 / A10G | — | Nvidia A10 $1.1016/hr Preemptible $804.17/mo Non-preemptible $3.3048/hr · $2,413/mo checked November 1, 2026 | — | |||
| RTX 4090 | — | — | — | — | ||
| A100 | Nvidia A100 80GB $2.4984/hr Preemptible $1,824/mo Non-preemptible $7.4952/hr · $5,471/mo checked November 1, 2026 | |||||
| H100 | Nvidia H100 SXM5 $3.9492/hr Preemptible $2,883/mo Non-preemptible $11.85/hr · $8,649/mo checked November 1, 2026 |
Prices checked November 1, 2026. A month is 730 hours. Tokens a month = documents × average tokens × re-embeds a month. The bill is GPU rental only. The pick is the highest MMTEB score among setups whose bill is no more than twice the cheapest scored setup, or $10 more, whichever is larger. Models without a score are priced but not picked.
Embedding is a forward pass over every token, so it is compute-bound. Rows without a published figure take the nearest published number for that model, or for a reference model of the same class, and multiply by (reference parameters ÷ model parameters) × (GPU dense BF16 TFLOPS ÷ the reference GPU's). Models of 4B parameters and up use E5-Mistral-7B, measured with vLLM. Models under 1B use BGE-M3's approximate TEI figure. Those rows are marked estimates. Real throughput depends on sequence length and batch size, and very small models usually hit the CPU tokenizer first, so their estimates run high.
Related tools: the self-hosted OCR cost calculator, the speech-to-text cost calculator, and the open-source LLM hosting cost calculator. Related reading: LLM API pricing, LLM cost optimization, and AI agent development.
For one-off batch jobs the GPU time is usually small, and text-embedding-3-small is also small: $0.02 per million tokens. The measured vLLM run in this tool embedded 36,600 tokens a second with E5-Mistral-7B on one H100, about 132 million tokens an hour. Self-hosting tends to win when you re-embed often, need a model the APIs do not offer, or cannot send the text out. An endpoint left on all month is billed for 730 hours whether it is busy or not.
Of the models in this tool, Qwen3-Embedding-8B has the highest MTEB Multilingual mean in the Qwen3-Embedding model card's leaderboard table: 70.58, then Qwen3-Embedding-4B at 69.45 and Qwen3-Embedding-0.6B at 64.33. BGE-M3 is 59.56 in the same table. Scores move as the leaderboard is updated, so check the live board before you choose.
They are not in the one table this page uses, and mixing scores from different MTEB versions would make the column misleading. A dash means no score from that table, not a low score. Their own model cards publish other numbers.
Two sources publish numbers: a measured vLLM run of E5-Mistral-7B (36.6k tokens a second on an H100, 15.3k on an A100) and a Spheron blog with approximate TEI figures for BGE-M3 and Qwen3-Embedding-0.6B. Everything else scales a reference model by parameter count and the GPU's dense BF16 TFLOPS, and is marked an estimate.
Not under its open licence. The Hugging Face card lists CC-BY-NC-4.0, which is non-commercial. The calculator still prices it, and the table flags the licence.
It multiplies the corpus. One million documents of 500 tokens is 500 million tokens each time you embed them. Weekly re-embedding is about 2.2 billion tokens a month. Most corpora only need new and changed documents embedded, so if you can track changes, enter that count instead.