deepinfra: Add script & sync latest DeepInfra models

[GodTamIt]

DeepInfra Model Sync Changes

Summary

This change adds a new script to keep DeepInfra models in sync. It then uses the script to update the current corpus of models from DeepInfra updated.

Note: The script uses an allowlist approach for providers - only explicitly approved providers are included. This is intentional as most models from DeepInfra should not be tracked in the database.

---

Changes Overview

DeepInfra Sync Script

Added: packages/core/script/generate-deepinfra.ts

Script Features

Data Fetching:

• Connects to https://api.deepinfra.com/v1/openai/models endpoint
• Fetches model metadata including pricing, context limits, and capabilities

Smart Filtering:

• Provider allowlist - only includes approved providers (strict by design)
• Regex filters skip embedding models, FLUX models, and image-specific models

CLI Flags:

• --dry-run: Preview changes without writing files
• --new-only: Only create new models, skip updates
• --no-delete: Keep orphaned files instead of deleting

---

Sync DeepInfra Models

New Models Added

Qwen (Alibaba):

• Qwen3-14B, Qwen3-32B
• Qwen3-235B-A22B-Instruct-2507 (with MoE architecture)
• Qwen3-235B-A22B-Thinking-2507 (reasoning variant)
• Qwen3-30B-A3B (MoE variant)
• Qwen3-Max and Qwen3-Max-Thinking
• Qwen3-Next-80B-A3B-Instruct
• Qwen3-VL-235B-A22B-Instruct (vision-language)
• Qwen3-VL-30B-A3B-Instruct
• Qwen3.5 series: 0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B

DeepSeek:

• DeepSeek-OCR - OCR-specialized model
• DeepSeek-R1-0528-Turbo - Optimized reasoning
• DeepSeek-R1-Distill-Llama-70B - Distilled variant

Anthropic:

• claude-4-sonnet - New Sonnet model supported

Meta Llama:

• Llama-3.2-11B-Vision-Instruct - Vision-capable model
• Llama-Guard-4-12B - Safety model

OpenAI:

• gpt-oss-120b-Turbo - Optimized variant

NVIDIA:

• Llama-3.3-Nemotron-Super-49B-v1.5
• NVIDIA-Nemotron-3-Super-120B-A12B
• NVIDIA-Nemotron-Nano-12B-v2-VL (vision-language)
• NVIDIA-Nemotron-Nano-9B-v2
• Nemotron-3-Nano-30B-A3B

Moonshot AI:

• Kimi-K2.5-Turbo - Optimized variant

StepFun:

• Step-3.5-Flash

Models Updated (22)

Capability & Pricing Updates:

• MiniMax-M2.1 and MiniMax-M2.5 - Updated capabilities
• Qwen3-Coder-480B-A35B-Instruct and -Turbo - Fixed pricing fields
• claude-3-7-sonnet-latest - Updated metadata
• claude-4-opus - Updated capabilities
• DeepSeek-R1-0528 - Revised pricing structure
• DeepSeek-V3.2 - Updated context limits
• Llama-3.3-70B-Instruct-Turbo - Capability adjustments
• Llama-4-Maverick-17B-128E-Instruct-FP8 - Metadata updates
• Llama-4-Scout-17B-16E-Instruct - Capability refinements
• Kimi-K2-Instruct-0905, Kimi-K2-Thinking, Kimi-K2.5 - Updated pricing
• gpt-oss-120b and gpt-oss-20b - Pricing adjustments

Zhipu AI (zai-org) GLM Series:

• GLM-4.6, GLM-4.6V, GLM-4.7, GLM-4.7-Flash, GLM-5 - Comprehensive updates to pricing, context limits, and capabilities

Models Removed (2)

Obsolete Models:

• MiniMaxAI/MiniMax-M2 - No longer available via API
• zai-org/GLM-4.5 - Discontinued

---

How to Use

Running the Sync

# Preview changes without writing files
bun run packages/core/script/generate-deepinfra.ts --dry-run

# Sync only new models (preserve existing manual edits)
bun run packages/core/script/generate-deepinfra.ts --new-only

# Full sync (add new, update existing, remove obsolete)
bun run packages/core/script/generate-deepinfra.ts

Validation

After running the sync, validate the generated configurations:

bun validate

---

Impact

This change significantly expands the DeepInfra provider coverage, adding:

• New models from major providers (Anthropic, Meta, Qwen, DeepSeek)
• Updated pricing reflecting current DeepInfra rates
• Consistent TOML formatting across all model files

Fixes #1043

[GodTamIt]

I filtered out a lot of unnecessary or older models in the change so it doesn't overwhelm the menus in OpenCode, per some discussion in Discord.