v0.58.0

QAI Hub Models v0.58.0 — 2026-07-15

Changelog

New Models

• Qwen3-VL-8B-Instruct (qwen3_vl_8b_instruct) — Qualcomm-optimized build of Qwen3-VL-8B-Instruct, a vision-language model, published and available on AI Hub.
• Qwen3-1.7B (qwen3_1_7b, QAIRT recipe) — added a QAIRT deployment recipe using a SpinQuant quantization flow.

New LLM / VLM runtime support

• Added geniex_llamacpp support for Qwen3-4B-Instruct-2507.
• Memory-mapped prefilled inputs for LLMs to reduce host memory pressure during eval.

Model improvements & fixes

• YOLO post-processing — three performance fixes across the YOLO family.
• CVT — added w8a16_mixed_int16 precision support.
• hrnet_pose — corrected the evaluation dataset.
• Qwen tokenizer/checkpoint — fixed the resolve_default_checkpoint stub and tokenizer handling for newer transformers releases.
• Voice AI SDK links — MeloTTS and Opus-MT now point to the correct Voice AI SDK for their task: MeloTTS → Voice AI TTS, Opus-MT → Voice AI Translation (Whisper continues to link Voice AI ASR).
• export.py now defaults to the model's default runtime; fixed component input-shape argument handling.

Deprecations

• The legacy Genie runtime remains deprecated in favor of GenieX (a deprecation warning is now surfaced). Pre-exported Genie assets and perf/accuracy data will not be updated in future releases.

Other

• Robotics tag added and applied to Pi-0.5 and GR00T-N1.5.
• Export/evaluate pipeline refactored into shared utils.export utilities.
• Dedicated GPU device pools for cs_8_elite_qrd / cs_x_elite to improve scorecard reliability.

SDK versions

• No SDK version changes in this release (default QAIRT remains 2.45).