embedding model comparison
Qwen3 Embedding 8B vs BGE-M3
BGE-M3 costs less than Qwen3 Embedding 8B — $0.012 against $0.030 per 1M tokens — so at a 500M-token corpus it comes to $6.00 against $15.00, a difference of $9.00, and the cheaper of the two is the one you cannot call yet. BGE-M3 is catalogued but not currently servable — a request naming `baai/bge-m3` returns `no available channel` — so its figure is what it will cost once an upstream carries it, while Qwen3 Embedding 8B can be called today at the rate above.
BGE-M3 is not available yet. No configured upstream serves BGE-M3, so a request returns no available channel. The rates below are what it will cost once an upstream carries it, not something you can spend today. Qwen3 Embedding 8B can be called now.
Side by side
| Qwen3 Embedding 8B | BGE-M3 | |
|---|---|---|
| Our rate — input | $0.030/M | $0.012/M |
| Official rate — input | $0.050/M | $0.020/M |
| Below official | 40% | 40% |
| Context window | 32K | 8K |
| Vector dimensions | 4,096 | 1,024 |
| Publisher | Alibaba | BAAI |
| Licence | Apache 2.0 | MIT |
| Released | Jun 5, 2025 | Jan 30, 2024 |
| Model id | qwen/qwen3-embedding-8b | baai/bge-m3 |
| Callable today | Yes | No |
What it costs, at a 500M-token corpus
One volume, both models, this catalogue’s published rates. Your figure will differ; the arithmetic will not, and the ratio of the two columns is the part that transfers to your own numbers.
| Model | At our rates | At official rates |
|---|---|---|
| Qwen3 Embedding 8B | $15.00 | $25.00 |
| BGE-M3Not servable | $6.00 | $10.00 |
At this volume BGE-M3 is $9.00 cheaper than Qwen3 Embedding 8B — 60% less. That ordering holds at any ratio of input to output, because the cheaper model is cheaper on every rate rather than on one column.
Excludes cached-input savings, which on a repeated-prefix workload reduce the input side substantially. Cached rates are in the table above and on the pricing page.
When to pick each
Pick Qwen3 Embedding 8B
- It holds 32K in one request against 8K — 4.0× wider, which decides whether a long input is one call or a map-reduce.
- Vectors are 4,096 dimensions against 1,024, which sets storage and search cost for as long as the index exists.
- It is published by Alibaba under the Apache 2.0 licence, where BGE-M3 is MIT.
- It is callable today. BGE-M3 is catalogued but not currently served by any upstream.
- The catalogue lists it for retrieval-augmented generation over large corpora, multilingual semantic search and deduplication and clustering at scale.
Pick BGE-M3Not servable
- It is not callable today — no configured upstream serves it, so the rate above is what it will cost rather than what you can spend.
- It bills $0.012 against Qwen3 Embedding 8B's $0.030 per 1M tokens — the only rate either model charges, so it is the whole of the price comparison.
- Vectors are 1,024 dimensions against 4,096, which sets storage and search cost for as long as the index exists.
- It is published by BAAI under the MIT licence, where Qwen3 Embedding 8B is Apache 2.0.
- The catalogue lists it for general-purpose vector search on a budget, long-document chunk embedding up to 8K tokens and self-hostable-later retrieval stacks.
Every line above is a field in the model catalogue rather than a judgement about output quality. This site publishes no benchmark results, so which model answers your prompts better is a measurement to run on your own traffic — the same key calls both.
Questions
- Which is cheaper, Qwen3 Embedding 8B or BGE-M3?
- BGE-M3, at a 500M-token corpus: $6.00 against $15.00, a difference of $9.00 — 60% less. That holds at any ratio of input to output on this pairing, because BGE-M3 has the lower rate on every column, not just the one the headline quotes. What the arithmetic does not tell you is whether both models do your job to the same standard; that is a measurement on your own prompts, not a number we can publish.
- Can I call both Qwen3 Embedding 8B and BGE-M3 today?
- Only Qwen3 Embedding 8B. A request naming `qwen/qwen3-embedding-8b` reaches a live upstream and is billed at the rate above; `baai/bge-m3` returns `no available channel`, because no configured upstream carries BGE-M3 yet. Its rate is published because it is what the model will cost, not because you can spend it today. 7 of the 24 models in this catalogue are callable right now; the rest are listed at their intended price and marked, so you find out here rather than from an error at call time.
- How much context does Qwen3 Embedding 8B hold compared with BGE-M3?
- Qwen3 Embedding 8B takes 32,768 tokens (32K) in a single request against BGE-M3's 8,192 (8K) — 4.0× wider. The window is a cost decision before it is a capability one: an input that does not fit has to be split, summarised in pieces and summarised again, which bills the source text more than once. Below that line the wider window is capacity you are not using.
- What licences are Qwen3 Embedding 8B and BGE-M3 under?
- Qwen3 Embedding 8B is published by Alibaba under Apache 2.0 and BGE-M3 by BAAI under MIT. Where one of those is a plain MIT or Apache 2.0 text and the other is a publisher's own or a modified variant, the difference is real and lives in the document rather than in the name — user thresholds, field-of-use terms and attribution requirements all appear in licences of that kind. Each model's own page carries the longer answer, and the licence text carries the binding one.
- How do I switch between Qwen3 Embedding 8B and BGE-M3?
- Change the model id and nothing else. Both are served from `https://router.xark.io/api/v1` behind the same OpenAI-compatible contract — `qwen/qwen3-embedding-8b` and `baai/bge-m3` — so the request shape, the streaming frames, the usage block and the error envelope are identical between them. Send an explicit id and no substitution happens underneath you, which is what makes running both and comparing the two on your own traffic a change of one string rather than a migration. BGE-M3 is not servable yet, so that swap is not available today.
- How do these rates compare with the official ones?
- Both rows are published side by side rather than as a single discount claim. Qwen3 Embedding 8B is 40% below Alibaba's official rate and BGE-M3 is 40% below BAAI's, which at a 500M-token corpus is $15.00 against an official $25.00 for Qwen3 Embedding 8B, and $6.00 against $10.00 for BGE-M3. The claim on this site is only that our rate sits below the model publisher's own — not that no one else is cheaper. Where a third-party provider undercuts us on either model, that provider's rate is printed on that model's own page.
Other comparisons: · all comparisons