Add three NVIDIA NIM models: Qwen3.5-397B, MiniMax-M2.5, and Step-3.5-Flash

[lcwecker]

Summary

Add three new model configurations for NVIDIA NIM provider, verified against official NVIDIA NIM documentation.

Changes

1. Qwen3.5-397B-A17B (providers/nvidia/models/qwen/qwen3.5-397b-a17b.toml)

• Type: 397B MoE multimodal model (text, image, video)
• Context: 262,144 tokens (native), extensible to 1M with YaRN
• Output: 81,920 tokens
• Release Date: 2026-02-16
• Source: https://build.nvidia.com/qwen/qwen3.5-397b-a17b

2. MiniMax-M2.5 (providers/nvidia/models/minimaxai/minimax-m2.5.toml)

• Type: 230B parameter coding and reasoning model
• Context: 204,800 tokens
• Performance: SOTA on SWE-Bench Verified (80.2%)
• Release Date: 2026-02-12
• Source: https://build.nvidia.com/minimaxai/minimax-m2.5

3. Step-3.5-Flash (providers/nvidia/models/stepfun-ai/step-3.5-flash.toml)

• Type: 196B MoE reasoning model (~11B active parameters)
• Context: 256,000 tokens
• Throughput: 100-300 tok/s
• Release Date: 2026-02
• Source: https://build.nvidia.com/stepfun-ai/step-3.5-flash

Verification

• All configurations validated with bun validate
• Specifications verified against NVIDIA NIM official documentation
• Model cards referenced from official sources

Notes

• Model specifications (context length, modalities, release dates) sourced from official NVIDIA NIM documentation
• Knowledge cutoff dates are estimates based on release dates where not explicitly stated by vendors