Skip to content

recommended-models.md

Latest commit

 

History

History
102 lines (78 loc) · 6.59 KB

recommended-models.md

File metadata and controls

102 lines (78 loc) · 6.59 KB
Raw

Recommended Models

Cheapest vision-capable models suitable for FindRomCover's AI Assist (synchronous chat completions, image input, text/JSON output).

Source: OpenRouter GET /api/v1/models — 446 models total, 275 accept image input (checked 2026-09-20). Prices are USD per million tokens (input / output), ordered by input price. FindRomCover's current OpenRouter default is google/gemma-4-26b-a4b-it.

Warning: This page is a snapshot of an external catalog dated 2026-09-20. Prices, model availability, context windows, and free tiers change frequently and without notice. Treat the figures below as guidance only and verify current details on the provider's website before running large batches.

Models

# Model ID Input $/M Output $/M Context JSON mode Notes
1 qwen/qwen3.7-flash $0.03 $0.13 1M response_format Cheapest overall; vision-language reasoning model from Alibaba; ~$0.21 per 1,000 picks
2 google/gemma-3-4b-it $0.05 $0.10 131k response_format Cheapest output tokens; small but genuinely multimodal
3 google/gemma-3-12b-it $0.05 $0.15 131k response_format Same input price as 4B with better quality
4 openai/gpt-5-nano $0.05 $0.40 400k structured outputs + response_format Most reliable JSON parsing; smallest GPT-5 variant
5 inclusionai/ling-3.0-flash-vl $0.06 $0.18 131k response_format Vision-language specialist (124B MoE / 5.5B active)
6 amazon/nova-lite-v1 $0.06 $0.24 300k prompt-only JSON Proven for cover picking — very reliable in real batch runs; solid low-cost multimodal; 5k max output tokens
7 qwen/qwen3.5-flash-02-23 $0.065 $0.26 1M response_format 1M context
8 ~z-ai/glm-flash-latest $0.075 $0.25 1.3M response_format 1.3M context
9 bytedance-seed/seed-1.6-flash $0.075 $0.30 262k response_format 262k context
10 google/gemma-3-27b-it $0.08 $0.45 131k response_format Larger Gemma 3
11 mistralai/mistral-small-3.2-24b-instruct $0.09 $0.25 256k response_format 256k context
12 google/gemma-4-26b-a4b-it $0.09 $0.30 262k response_format MoE variant of Gemma 4; FindRomCover default
13 google/gemma-4-31b-it $0.09 $0.34 262k response_format Best value of the Gemma 4 family
14 qwen/qwen3.5-9b $0.10 $0.15 262k response_format 262k context; text, image and video input
15 meta-llama/llama-4-scout $0.10 $0.30 1.3M response_format 1.3M context
16 google/gemini-2.5-flash-lite $0.10 $0.40 1M response_format Known-good quality; per-image pricing
17 qwen/qwen3.8-27b $0.20 $2.50 1M response_format Large Qwen vision model
18 thinkingmachines/inkling-small $0.45 $1.20 1.05M prompt-only JSON Multimodal, also accepts audio
19 thinkingmachines/inkling $1.00 $4.05 1.05M prompt-only JSON Multimodal, also accepts audio

Estimated cost per 1,000 cover picks (assuming about 6 images ≈ 6k prompt tokens + about 200 completion tokens):

Model Estimated cost per 1,000 picks
qwen/qwen3.7-flash ≈ $0.21
google/gemma-3-4b-it ≈ $0.32
google/gemma-3-12b-it ≈ $0.33
openai/gpt-5-nano ≈ $0.38
amazon/nova-lite-v1 ≈ $0.41

The default google/gemma-4-26b-a4b-it costs about $0.60 per 1,000 picks. See the table above for alternatives.

OpenCode Zen (free MiMo-V2.5)

OpenCode Zen is the OpenCode team's AI gateway. It is OpenAI-compatible and currently offers MiMo-V2.5 Free (mimo-v2.5-free) at no cost for a limited time. MiMo-V2.5 accepts image input, so it works with FindRomCover.

Setting Value
Provider Custom (OpenAI-compatible)
Base URL https://opencode.ai/zen/v1
API Key Your Zen key from opencode.ai/auth
Model mimo-v2.5-free

Notes:

  • Free for a limited time; Zen states that data from MiMo-V2.5 Free may be used to improve the model while it is free.
  • Zen's model metadata may not report image modality, so the picker's Vision-capable only filter can hide it — uncheck the filter or type the model ID directly.
  • Zen also serves paid vision models, for example deepseek-v4-flash-vision-exp ($0.14 / $0.28 per M tokens).
  • In OpenCode itself the model is referenced as opencode/mimo-v2.5-free; FindRomCover uses the raw model ID mimo-v2.5-free.

Caveats

  • :free variants exist for several models above but are rate-limited and frequently unavailable — use the paid endpoints for batch runs.
  • :batch variants are cheaper but asynchronous and do not work with FindRomCover's synchronous client.
  • Models without native JSON mode still work: FindRomCover prompts for JSON and parses it from the response text.
  • Image token cost is provider-dependent; the estimates above assume about 1,000 tokens per 512px thumbnail.
  • Reasoning models may spend output tokens on internal reasoning before producing text. FindRomCover allows up to 4,096 output tokens and reports a clear error if a model still exhausts the budget.

How to use

  1. Open Settings > AI Settings....
  2. Select OpenRouter and paste your API key.
  3. Click Test / Load Models.
  4. Type a model ID from the table into the filter box and select it.
  5. Save, then use AI Pick Best or Batch Fill.

All models listed here are flagged vision-capable by OpenRouter's modality metadata, so they appear when Vision-capable only is checked.

For OpenCode Zen, choose Provider: Custom (OpenAI-compatible), set the Base URL to https://opencode.ai/zen/v1, paste your Zen key, and enter mimo-v2.5-free as the model.

Choosing a model

Priority Recommendation
Lowest cost qwen/qwen3.7-flash
Most reliable JSON output openai/gpt-5-nano
Best quality per dollar google/gemma-3-12b-it or google/gemini-2.5-flash-lite
Best Gemma 4 value google/gemma-4-31b-it
Proven for cover picking amazon/nova-lite-v1 — very reliable in real batch runs
Free (OpenCode Zen, limited time) mimo-v2.5-free
No cloud cost A local model via Ollama, for example qwen2.5vl:7b

Note: Prices change frequently. Verify current pricing on your provider's website before running large batches.

Related pages