Vision on free tiers

Free multimodal is thinner than free chat, but real. From the catalog (pnpm mh list --capability vision --free-only):

Provider	Model	Context	Free limits
Google AI Studio	gemini-2.5-flash	1M	10 RPM class (see guide); training caveat applies
Groq	meta-llama/llama-4-scout-17b-16e-instruct	131K	30 RPM / 1,000 RPD
Cloudflare Workers AI	@cf/meta/llama-4-scout-17b-16e-instruct	—	10,000 neurons/day
OVHcloud	Qwen2.5-VL-72B-Instruct	—	anonymous ~2 RPM/IP; no-training policy
Cerebras	gemma-4-31b	—	official free tier (tight RPM)
Z AI	glm-4.6v-flash	—	permanent free, 1 concurrent request

Wire it

pnpm mh recommend --capability vision --snippet

OpenAI-style image input through the gateway:

const res = await client.chat.completions.create({
  model: "auto",
  messages: [
    {
      role: "user",
      content: [
        { type: "text", text: "What's in this image?" },
        { type: "image_url", image_url: { url: "data:image/png;base64,..." } },
      ],
    },
  ],
});

Caveat: "OpenAI-compatible" vision payloads diverge more than chat payloads (base64 vs URL support, size caps). The gateway passes your payload through untranslated — test each rung once (pnpm mh probe --live --provider <id>) before trusting a cascade with images.