Guides

Open weights, one field away, cheaper past 2K tokens.

Same key, same base URL. Darkbloom serves quantized open-weight models on Apple Silicon, with a floor on every request.

Get a key — no account
Base URLhttps://api.minirouter.sh/v1

At a glance

Models on Darkbloom
63 only here
Best input rate
7.5× lowerQwen 3.6 35B A3B
Floor per request
$0.000105When tokens cost less
Precision
2-bit / 4-bit / 8-bitNamed per model

One line in. One header back.

the request · chat completions or the Responses API
curl https://api.minirouter.sh/v1/chat/completions \  -H "Authorization: Bearer $MINIROUTER_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "openai/gpt-oss-20b",    "messages": [{"role": "user", "content": "ping"}],added line:     "providerOptions": {"gateway": {"only": ["darkbloom"]}}  }'
the response
HTTP/2 200added line: x-minirouter-provider: openai/gpt-oss-20b@darkbloomx-minirouter-cost-usd: 0.000105x-minirouter-request-id: req_…

Pinned means pinned. Unpinned means unchanged.

Where a request goesA request with providerOptions.gateway.only set to darkbloom is served on Darkbloom, or refused with 503 and nothing charged. A request without the field goes to the default route when the model has one, and to Darkbloom when only Darkbloom serves the model.servedunavailabledefault route existsDarkbloom onlyYour requestPOST /chat/completionsField setonly: ["darkbloom"]No fieldrequest unchangedDarkbloom@darkbloomRefused503, nothing chargedDefault route@vercelDarkbloom@darkbloom

Per million tokens, side by side

After markup, from the live catalog. The default route is what an unpinned request pays.

6 models with a Darkbloom route, current catalog rates
ModelInput, $ per 1M tokensOutput, $ per 1M tokens
Qwen3 VL 30B A3B Instruct4-bit on Darkbloom
$0.095
—
Darkbloom only
$0.42
—
Darkbloom only
Qwen 3.5 9B4-bit on Darkbloom
$0.084
—
Darkbloom only
$0.137
—
Darkbloom only
Qwen 3.6 35B A3B4-bit on Darkbloom
$0.053
$0.394
7.5× lower on Darkbloom
$0.735
$2.363
3.2× lower on Darkbloom
Google Gemma 4 26B A4B4-bit on Darkbloom
$0.044
$0.158
3.6× lower on Darkbloom
$0.231
$0.63
2.7× lower on Darkbloom
GPT OSS 20B8-bit on Darkbloom
$0.021
$0.032
1.5× lower on Darkbloom
$0.105
$0.147
1.4× lower on Darkbloom
Bonsai 2 27B2-bit on Darkbloom
$0.079
—
Darkbloom only
$0.525
—
Darkbloom only
DarkbloomDefault routeOne scale across both columns.

Cheaper past 2K input tokens

GPT OSS 20B, reply held at 250 tokens. Below the crossing the default route wins; above it Darkbloom does, and the gap widens.

Cost per request on GPT OSS 20B, 250 output tokens, by prompt sizeCost per request on GPT OSS 20B as the prompt grows from 0 to 8K input tokens with 250 output tokens. Darkbloom is flat at its $0.000105 floor until about 2K input tokens, where the default route costs the same; beyond that Darkbloom is cheaper.$0.000078$0.000156$0.000234$0.00031202K4K6K8Kinput tokens per request · output fixed at 250floor $0.000105cheaper past 2K tokensdefault routeDarkbloom
Cost per request on GPT OSS 20B, 250 output tokens, by prompt size, as a table
Input tokensDarkbloomDefault routeCheaper
0$0.000105$0.000037Default
250$0.000105$0.000045Default
500$0.000105$0.000053Default
1K$0.000105$0.000068Default
2K$0.000105$0.000100Default
4K$0.000110$0.000163Darkbloom
8K$0.000194$0.000289Darkbloom

Four workloads on GPT OSS 20B

Health check
A ping. · 20 in, 5 out
Darkbloom $0.000105 at the floor, default $0.000001
+7592%
Chat turn
Short history, short reply. · 600 in, 150 out
Darkbloom $0.000105 at the floor, default $0.000041
+156%
Agent turn
System prompt, tools, history. · 4K in, 250 out
Darkbloom $0.000110, default $0.000163
−32%
Document pass
A long file in, a summary out. · 30K in, 500 out
Darkbloom $0.000682, default $0.0010
−33%

Pin the real work; leave health checks and empty pings on the default route.

Keep going