# 0.12.0

## Release Information

- OS & Framework support - Tested with the following configurations:

> 
> 
> - Non-Gen AI:
> 
> 
> 
> > 
> > 
> > - Frameworks: ONNX, TFLite, Pytorch
> >         - Host Platforms: Linux-x86\_64 (Ubuntu 22.04, Ubuntu 24.04), Linux-aarch64 (Ubuntu 24.04, Python 3.12), Windows-x86\_64 (10+), Windows-arm64 (10+)
> >         - Target Platforms: Linux-x86\_64 (Ubuntu 22.04, Ubuntu 24.04), Linux-aarch64 (Ubuntu 24.04), Windows-arm64 (10+), Android-arm64, QNX
>     - Gen AI:
> 
> 
> 
> > 
> > 
> > - Frameworks: ONNX, GGUF
> >         - Host Platforms: Linux-x86\_64 (Ubuntu 22.04, Ubuntu 24.04)
> >         - Target Platforms: Android-arm64, Linux-aarch64
- QAIRT SDK - Tested with version (2.46.0, 2.50.0 (Default), 2.51.0)

Note

For QAIRT SDK, other supported SDK versions can be listed using `qairt-vm fetch --list`.

## Highlights

- Default QAIRT SDK updated to 2.50.0.
- New features in QAIRT Dev Python APIs (compatible with QAIRT SDK 2.50.0+, unless noted otherwise)

> 
> 
> - Gen AI Builder
> 
> 
> 
> > 
> > 
> > - `GenAIBuilderCPU.set_precision()` lets callers select a CPU export precision through
> > the new `Precision` enum, and `QuantizationLevel` gains `Q8`/`Q5_K` values,
> > extending CPU-backend Gen AI Builder exports beyond the previous fixed-precision path.
>     - Pipeline
> 
> 
> 
> > 
> > 
> > - The Pipeline API now supports end-to-end Qwen3.5-VL (vision-language) checkpoints,
> > loadable via `QcAutoModelForImageTextToText`, with basic and advanced VLM recipes
> > covering quantization and on-device generation.
> >         - LoRA quantization now uses a two-tier on-disk cache: Tier 1 caches the base-model
> > quantization for reuse across adapters, and Tier 2 caches per-adapter results, so
> > switching adapters no longer re-quantizes the shared base model.
> >         - The Pipeline API adds AIMET HF SpinQuant enablement (`spinquant` /
> > `spinquant_adascale` recipes) with R1+R2 Hadamard rotations for improved
> > post-training quantization accuracy.
> >         - Eaglet v2 draft models can now be evaluated for `Block Efficiency` (BE) — the
> > average number of tokens accepted per speculation step — via a new
> > `evaluator_config` metric driven by `eval_tm_arn`/`eval_dm_arn` and
> > `pipe.evaluate()`, letting draft-model quality be validated before compiling the
> > on-device container.
> >         - `Pipeline.construct()` now renders a live terminal progress display (per-stage
> > and overall), which auto-disables on non-interactive stdout and can be turned off
> > explicitly with `show_progress=False`.
> >         - A new `evaluation_mode` recipe flag evaluates every stage inline as
> > `construct()` runs, logging each stage’s score and a Δ% vs. FP32 summary table
> > without a separate `pipe.evaluate()` call — useful for debugging accuracy
> > regressions as they’re introduced stage by stage.

## Resolved Issues

> 
> 
> - Fixed an on-device accuracy regression for Qwen3.5 text models: `LLMSplitter` now captures
> multi-node terminal output branches within split boundaries and registers dispatch for
> `Qwen3_5ForCausalLM`.
> - Fixed Gen AI export and cache workflows writing temporary files to a hardcoded or system temp
> directory instead of honoring `QAIRT_TMP_DIR`, which could leave artifacts outside the
> configured workspace or fail in environments with a restricted system temp directory.

## Deprecation Notices

To be removed in 0.14.0:

> 
> 
> - `GenAIConfig.allow_async_init` and `GenAIConfig.enable_graph_switching` are deprecated. Set
> these through `EngineConfig.htp.allow_async_init` and `EngineConfig.htp.enable_graph_switching`
> and pass the `EngineConfig` to `get_executor()` / `T2TExecutor` instead.
> - `GenAIExecutor` (`qairt.gen_ai_api.executors.gen_ai_executor`) is deprecated and has been
> renamed to `GenAIExecutable` (`qairt.gen_ai_api.executors.gen_ai_executable`). Update imports
> to the new module.

## Known Issues

> 
> 
> - Gen AI Builder is temporarily unsupported on Windows-arm64. Use Linux-x86\_64 as the host platform
> for Gen AI workflows.
> - Windows on ARM64 is not functional. This is a known issue.
> - Models with batch size greater than 1
> 
> 
> 
> > 
> > 
> > - A discrepancy exists between the qnn-net-run CLI and the Python API regarding data preparation
> > for batch sizes greater than 1. While the CLI automatically groups individual tensor paths from
> > an input list into a batch, the Python API requires inputs to be pre-batched (concatenated into
> > a single raw file) prior to execution

Last Published: Oct 05, 2026

[Previous Topic
Known Issues](https://docs.qualcomm.com/bundle/publicresource/80-87189-2/topics/0_11_0.md)