Ternlight 7 MB Embedding Model: Pricing Impact
Ternlight is a 7 MB WASM embedding model that runs on-device. Here is the pricing impact for semantic search, RAG, and hosted embedding bills.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
Ternlight landed on Hacker News today with a small but useful pricing signal: the project ships a 7 MB WebAssembly embedding model that can run in the browser, Node, edge runtimes, and small devices without calling a hosted API.
This is not a frontier model launch, and it is not a new hosted API rate card. The pricing impact is narrower: embedding workloads are one of the easiest parts of an AI stack to move local when quality requirements are modest. If a 7 MB model is good enough for search boxes, FAQ matching, intent routing, product-card similarity, and small RAG utilities, some teams can cut a recurring per-call API line item down to device CPU time and engineering maintenance.
Ternlight’s model card says the default shipped binary is 4.6 MB, the full WASM bundle is 7 MB, latency is roughly 2 ms per call on the tested M-series Mac setup, and the model is designed for short-string semantic similarity rather than long-document understanding. The project is MIT licensed and distributed through npm and Hugging Face.
For hosted alternatives, compare Cohere pricing, OpenAI pricing, and Mistral pricing, then model your actual request volume in the AI token cost calculator. For broader architecture tradeoffs, see our local AI vs API vs subscription pricing guide and cached tokens cost savings guide.
What Changed
The July 7 alert is the public availability of Ternlight as a tiny embedding model intended for local deployment. According to the model card, Ternlight is a 1.58-bit BitNet-style sentence embedding model distilled from sentence-transformers/all-MiniLM-L6-v2, with quantization-aware training and post-training int4 quantization.
The buyer-relevant details are:
| Detail | Ternlight claim |
|---|---|
| Full WASM bundle | 7 MB |
| Default shipped binary | 4.6 MB |
| Parameters | About 9.5M |
| Output dimension | 384 |
| Max input | 128 tokens |
| Reported p50 latency | About 2 ms |
| Reported throughput | About 450 embeddings/sec |
| License | MIT |
| Intended deployment | Browser, Node, edge runtimes, ARM boards |
The important caveat is the 128-token cap. Ternlight is not designed to embed entire documents in one call. It is closer to a local utility model for short queries, titles, tags, FAQ entries, product snippets, and routing labels.
Pricing Comparison
Hosted embedding APIs are already cheap on a per-token basis. Cohere Embed v3 English and Embed v3 Multilingual are listed in our live data at $0.10 per million input tokens with no output-token charge. That is inexpensive enough that most prototypes should still start with a hosted embedding API.
But at high volume, even cheap embedding calls become a real product cost. The local route changes the equation from per-token billing to bandwidth, CPU, memory, packaging, and quality-control work.
| Route | Listed token price | Best fit | Pricing read |
|---|---|---|---|
| Ternlight local WASM | No per-call API fee | Short semantic search, FAQ matching, on-device routing | Pay in engineering work and client/server CPU |
| Cohere Embed v3 English | $0.10 / 1M input tokens | Production embeddings with managed API reliability | Very low hosted baseline |
| Cohere Embed v3 Multilingual | $0.10 / 1M input tokens | Multilingual embedding pipelines | Ternlight is English-only, so this stays relevant |
| GPT-5.4 nano | $0.20 / 1M input, $1.25 / 1M output | Lightweight classification or generation | More flexible than embeddings, but not free at scale |
| Gemini 2.5 Flash-Lite | $0.10 / 1M input, $0.40 / 1M output | Low-cost synthesis and routing | Useful when you need generated answers |
| Command R7B | $0.0375 / 1M input, $0.15 / 1M output | Cheap RAG and classification | Often cheaper than using a frontier model for every step |
For a simple monthly workload of 500M short input tokens for search indexing and query embeddings, Cohere Embed v3 at $0.10/M would cost about $50/month before any storage, vector database, or reranking costs. At 5B tokens, the embedding line becomes about $500/month. At 50B tokens, it becomes $5,000/month.
Those numbers are still small compared with premium generation costs, but they are not zero. If the workload is simple, English-only, latency-sensitive, or privacy-sensitive, local embeddings become worth testing.
Who Benefits
Local-first app builders benefit first. A browser search feature, desktop notes app, offline documentation tool, or private FAQ assistant can embed short text without sending every query to a server. That can reduce latency, simplify privacy positioning, and avoid a metered embedding dependency.
RAG teams benefit when embeddings are part of a routing layer rather than the only quality-critical retrieval step. For example, a product can use a tiny local model to detect intent or narrow a candidate set, then send only harder cases to a hosted embedding API, reranker, or premium model.
Hosted API providers still benefit in production-grade use cases. Ternlight does not replace multilingual embeddings, long-text embedding models, managed uptime, monitoring, batch indexing, vector database integrations, or high-quality retrieval over messy corpora. It mainly raises the bar for what buyers should expect to run locally for simple tasks.
What Buyers Should Test
Do not move embeddings local only because the model file is small. Run an evaluation against your actual search or routing data.
Track:
| Metric | Why it matters |
|---|---|
| Top-k retrieval quality | Measures whether users find the right result |
| False positive matches | Local similarity can feel cheap until bad matches hurt trust |
| Language coverage | Ternlight is English-only |
| Query length | Inputs above 128 tokens are the wrong fit |
| Cold-start time | Important for browser and edge delivery |
| Bundle size impact | 7 MB is tiny for AI, but still large for some web apps |
| Fallback rate | Shows how often you still need hosted embedding or reranking |
The best architecture is often hybrid. Use local embeddings for low-risk, short-string matching; use Cohere, OpenAI, Google, or specialist embedding providers when retrieval quality, multilingual coverage, observability, or enterprise support matters.
Practical Advice
Start with a hosted embedding API if you are still validating product-market fit. At $0.10 per million input tokens, Cohere Embed v3 is cheap enough that engineering time will usually dominate the early cost model.
Test Ternlight when one of three constraints matters: privacy, offline use, or very high volume. A local model can be especially attractive for browser extensions, local documentation search, personal knowledge bases, point-of-sale apps, field tools, and internal utilities where short text is the main retrieval unit.
Do not use it as a drop-in replacement for every RAG pipeline. Long PDFs, legal archives, multilingual support centers, and semantic search over messy enterprise data should keep a stronger hosted embedding model or reranker in the evaluation set.
Finally, price the whole workflow. Embeddings are only one line item. Vector storage, reranking, answer generation, cache strategy, observability, and human review can dominate the actual bill. Local embeddings save money only when they preserve enough retrieval quality to reduce hosted calls without increasing support or review work.
Bottom Line
Ternlight is a useful reminder that not every AI pricing decision is about frontier token rates. Some workloads should not be metered at all if a small local model can do the job.
For buyers, the right takeaway is selective substitution: keep hosted embeddings for production quality and multilingual coverage, but benchmark local WASM embeddings for short, private, high-volume semantic tasks. If the quality holds, the price changes from cents per million tokens to a fixed engineering and compute cost.
Sources: Ternlight on Hugging Face, Ternlight GitHub repository, Hacker News discussion, and AI Pricing Guru’s live pricing dataset.