Cloud
Token billing. Ready today.
Call the same OpenAI-compatible API over the internet. Pay per million tokens. Competitive rates, no credit buckets, no usage theater.
Midium Cloud does not retain prompt data or generated response data.
Local
Air-gapped. On your Mac.
- Mac Mini$150/mo
- Mac Studio$300/mo
- Vaults add-on+$100/mo
Download the desktop app for out-of-the-box production-grade inference on hardware you own. Same models. Same API. Unlimited on the box.
Midium Local is 100% air-gapped — built for compliance and data-sensitive work. Nothing leaves the machine. Vaults are a +$100/mo add-on: no vault count or usage restrictions.
Tool calling
The engine that makes agents reliable.
Everyone serves the same open models. Midium’s reliability layer constrains generation so native OpenAI tool_calls stay well-formed — even when the raw model would fumble them. Same contract on cloud and local. Measured on Gemma 4 26B A4B against Ollama: identical 4-bit weights, Apple Silicon, BFCL v4’s official checker, 1,000 cases, temp 0.
- BFCL tool-calling accuracy
- 0.902
- vs Ollama’s best 0.874
- Parallel-call points
- +13.5
- Matched no-think · 0.830 vs 0.695
- Decode speed
- +37%
- 123.8 vs 90.1 tok/s
- Sooner to first token
- 2.3×
- 0.12s vs 0.28s
Parallel calls are the reliability moat. On the much smaller E4B with thinking on, Ollama edges Midium — the flagship 26B is where the layer shows. Courier vs Ollama: Tool-Calling, Measured.
Vaults
The database for your agents.
Persistent knowledge bases, attached over MCP. Retrieval is semantic and exact. Agents write back as they work.
Cloud
Unlimited vaults. Pay for traffic.
- Reads / 1k$5
- Writes / 1k$25
A read is an MCP retrieve, search, or get. A write is an insert, update, or delete that mutates the index.
Local
+$100/mo. No restrictions.
Add Vaults to a Mini or Studio license. Same MCP surface, air-gapped, unlimited vaults, no usage caps. The company brain stays on a machine you own.
- MCP-accessible — attach a vault the way you attach any tool.
- Auto-writeback — agents modify the vault as they work.
- Semantic and deterministic retrieval in one engine.
Local stays private
Prompts, files, outputs, and local vaults stay on your machine. The runtime never phones home with customer data. Air-gapped by architecture, not by policy.
Cloud keeps nothing
Hosted inference is billed by the token. We do not retain prompts or generated responses.
Vaults persist
Vault contents persist until you delete them — that is the product. We do not train on them. Local vaults never leave the machine.
Library
Models and costs
Cloud prices are per million tokens. Input is prefill. Output is decode. Local includes the full library on either a Mac Mini license ($150/mo) or a Mac Studio license ($300/mo).
| Model | Class | Cloud input / M | Cloud output / M | Local |
|---|---|---|---|---|
| Gemma 4 E2B | Lite | $0.01 | $0.05 | Included |
| Gemma 4 E4B | Lite | $0.014 | $0.07 | Included |
| Gemma 4 26B A4B | Balanced | $0.028 | $0.14 | Included |
| Qwen3.8 27B | Balanced | $0.032 | $0.16 | Included |
| Gemma 4 31B | Balanced | $0.054 | $0.27 | Included |
| Laguna XS 2.1 | Balanced | $0.052 | $0.26 | Included |
| Laguna S 2.1 | Frontier | $0.184 | $0.92 | Included |
| Inkling Small | Frontier | $0.30 | $1.50 | Included |
Same catalog on Midium Cloud and Midium Local. OpenAI-compatible API.
