
You built a small personal memory tool — a notes search, a chat-history RAG, a "second brain" that actually answers questions. It works. Then the embedding bill arrives, and it is absurdly large for a project only you use. A developer on LINUX DO ran into exactly this: their personal memory backend, built on the open-source tool Hindsight, was burning several yuan a day on Alibaba Cloud embeddings alone. The thread that followed turned into a useful survey of how Chinese developers keep embedding costs at or near zero, and the answers generalize well beyond China.
The short version: for personal-scale retrieval, you almost never need to pay per token. A free hosted tier covers most hobby projects, and the best open embedding models run comfortably on a laptop CPU.
Embedding pricing looks trivial on a rate card. Alibaba Cloud's Model Studio lists its text-embedding-v4 model at about $0.07 per million input tokens internationally (¥0.50 per million in Beijing), and OpenAI's small embedding model is around $0.02 per million. At those rates, a million tokens sounds like a lot of text. It is not, once a RAG pipeline is involved.
A personal memory tool re-embeds constantly: every note you save, every chat message you log, every document you import. Worse, naive pipelines re-embed whole documents when a single paragraph changes, and chunking multiplies token counts — a 100,000-word note archive becomes several hundred thousand tokens after overlap. Add a nightly re-index job and a "several yuan a day" bill stops being surprising.
For a personal project, work down this ladder and stop at the first rung that fits.
1. A free hosted embedding model. SiliconFlow, a Chinese API platform hosting open-weight models, marks several embedding and reranker models as free on its pricing page, including BAAI's bge-m3, bge-reranker-v2-m3, and the bge-large Chinese and English models. The API is OpenAI-compatible, so existing code usually needs only a base-URL change. Two caveats: the free catalog changes over time, so check the live pricing page before you build on a specific model ID; and third-party guides report that identity verification is required to use the free models, with fixed rate limits that cannot be raised by paying.
2. A platform credit allowance. ModelScope, Alibaba's model hub, runs an API-inference service where the Qwen3 embedding models are available behind a daily allowance of platform credits — one independent write-up measured roughly 200 credits granted per day for logging in, with each embedding call costing about one credit.
3. A vendor free trial quota. Alibaba Cloud's Model Studio gives each embedding model a one-million-token free trial quota, valid for 90 days. That is enough to prototype and to learn what your real usage looks like — which is the point. Measure first, then decide whether the paid rate, around $0.07 per million tokens, is actually a problem for your volume.
4. Your own machine. The open-weight models below run locally, which makes the marginal cost of one more embedding exactly zero. This is the rung the forum thread kept coming back to, and for a good reason: embedding models are small.
If you pick one model to self-host, BGE-M3 from the Beijing Academy of Artificial Intelligence (BAAI) is the consensus default, and the specifications explain why. It has about 568 million parameters, outputs 1024-dimensional dense vectors, handles up to 8,192 tokens of input, and supports more than 100 languages. It is released under the permissive MIT license, and the weights are freely available on Hugging Face.
What makes it unusual is that one checkpoint produces three kinds of embeddings at once: dense vectors for semantic search, sparse vectors for keyword-style lexical matching, and ColBERT-style multi-vector output for fine-grained late interaction. In practice, most people only use the dense output. But the sparse output is genuinely useful for personal notes, where exact terms — a person's name, a project codename, an error string — matter as much as meaning.
Alibaba's Qwen3 embedding family comes in three sizes — 0.6B, 4B, and 8B — and the 8B model topped the MTEB multilingual leaderboard with a score of 70.58 when it was released in June 2025. The family supports up to 32,768 tokens of context and up to 4,096 output dimensions, with Matryoshka-style dimension truncation that lets you store shorter vectors (say, 1024 dimensions) without re-embedding. The weights are Apache 2.0 licensed.
For personal use, the 0.6B model is the interesting one. It is roughly the same size class as BGE-M3, runs on CPU, and gives you the Qwen3 family's multilingual and code-retrieval strengths in a package your laptop can serve. The 8B model is measurably better on benchmarks, but it wants a real GPU — overkill for searching your own notes.
An embedding search returns the nearest candidates, but "nearest" is a blunt instrument — especially with short, ambiguous queries against a mixed personal corpus. A reranker takes the top 20 or 50 candidates and scores each one properly against the query, then reorders them. It is the single highest-leverage improvement you can make to a personal RAG pipeline, and it is cheap because it only runs on a handful of candidates.
The free options line up neatly with the embedding choices above. BGE's reranker, bge-reranker-v2-m3, is free on SiliconFlow alongside bge-m3, and the model itself is open-weight and self-hostable. Qwen3-Reranker-0.6B is the small local alternative. Either way, the pattern is the same: embed once, retrieve broadly, rerank narrowly.
| Stage | Free hosted option | Self-hosted option |
|---|---|---|
| Embedding | BAAI/bge-m3 on SiliconFlow | BGE-M3 or Qwen3-Embedding-0.6B |
| Reranking | bge-reranker-v2-m3 on SiliconFlow | bge-reranker-v2-m3 or Qwen3-Reranker-0.6B |
| Credit allowance | ModelScope API-Inference daily credits | — |
| Cheapest paid fallback | Bailian text-embedding-v4, ~$0.07/1M tokens | Electricity |
The fastest path to zero-cost embeddings is the hosted free tier, because it needs no local model download. With SiliconFlow's OpenAI-compatible endpoint, the embedding call looks like any other OpenAI-style request:
from openai import OpenAI
client = OpenAI(
base_url="https://api.siliconflow.cn/v1",
api_key="YOUR_SILICONFLOW_KEY",
)
resp = client.embeddings.create(
model="BAAI/bge-m3",
input=["Why is my Postgres replica lagging?", "Notes from the homelab rebuild"],
)
vectors = [d.embedding for d in resp.data]
Before depending on it, confirm on SiliconFlow's live pricing page that BAAI/bge-m3 is still marked free — the catalog changes, and a model that was free last quarter may not be this quarter.
The local route is barely longer. With sentence-transformers, BGE-M3 loads and embeds in a few lines, and nothing leaves your machine:
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("BAAI/bge-m3")
vectors = model.encode(
["Why is my Postgres replica lagging?", "Notes from the homelab rebuild"],
normalize_embeddings=True,
)
For a personal memory tool like Hindsight, which stores what you did and lets you search it later, the local route has a second advantage beyond price: your notes, messages, and browsing history never go to a third-party API at all. That privacy argument is arguably stronger than the cost argument, and it is the reason many developers in the thread preferred self-hosting even when a free API was available. If you want the full picture of what a local-first memory tool looks like, our earlier look at Hindsight as a local screen-time and memory tracker covers the setup.
Forum users also mentioned Gitee's free vector models and a community service at emrank.tech as further options. Both were reported in discussion rather than verified against primary sources, so treat them as leads to check, not recommendations — the pattern of this whole space is that free offers change faster than blog posts do.
The thread that started all this was not really about which embedding model is best. It was about a developer who discovered that a per-token meter, applied to a background job, produces a bill that feels wrong for the value received. The fix was never a cheaper model. It was matching the pricing shape to the workload shape: free tiers and local inference for always-on personal tools, paid APIs for bursts and for quality you cannot host.
That is also why the economics of personal AI tooling keep pointing the same way. The expensive part of a hobby RAG setup was never the model — it is the meter. Remove the meter, and the project that cost several yuan a day costs nothing, runs offline, and keeps your notes on your own disk. And if you are weighing the broader question of when to pay for API access at all, our breakdown of relay-station API economics covers the same metering question for chat models.