qwen3-asr-1.7b-glade

Qwen3-ASR 1.7B for Glade

Glade is a native Swift library for on-device speech inference on Apple silicon. This repository supplies converted model assets and configuration for Glade's model-specific host runtime. Glade's runtime source is currently private.

Glade conversion of Qwen3-ASR-1.7B by the Qwen Team, Alibaba Cloud, at source revision 7278e1e70fe206f11671096ffdd38061171dd6e5. The converted weights retain Apache-2.0 licensing; see LICENSE and NOTICE. This conversion is not endorsed by the upstream authors.

This repository contains one offline transcription bundle, approximately 2.68 GB, using 30-second chunks and up to eight independent decoder lanes. It prefers ANE for both the audio encoder and autoregressive text decoder. W8A16 quantization retains selected sensitive tensors in FP16.

Requires physical Apple silicon with macOS 27 or iOS 27 and Glade's model-specific host runtime. The .aimodel functions accept features, embeddings and KV caches, not audio files. The host must implement the frontend, prompt assembly, greedy generation and tokenization described by the metadata and sidecars. These assets do not load through Transformers or Core ML's MLModel API.

Bundle and semantics

• subsampling.aimodel and audio.aimodel preserve Qwen's frontend and native eight-second offline audio attention windows. Text attention sees all audio embeddings within each 30-second chunk.
• Four text_*.aimodel assets cover all 28 text layers. Each shares source weights across prefill and one-/eight-lane decode entrypoints.
• Compact KV caches start at 512 positions and expand to 1,024 when needed. Every valid context position is retained. Exhaustion is an error, not silent truncation. Independent chunks do not share decoder history.
• head.aimodel projects the vocabulary and returns winning partition/local IDs, preserving exact token IDs despite FP16 tensor transport.
• embeddings.f16, tokenizer.json, prompts.json, frontend tensors and metadata.json complete the runtime. Metadata owns capacities, stage boundaries, lane counts, language hints and EOS conventions.

The upstream model supports 30 languages and 22 Chinese dialects. The supplied prompts include its 30 language choices; automatic language selection is also possible by omitting the language hint. This conversion has English numerical and recording regression coverage, not a comprehensive evaluation of all languages and dialects.

No native word timestamps are produced. Applications that need them can run Qwen3 ForcedAligner as a separate, optional pass. Its download, preparation, memory and inference costs are excluded from the ASR measurements below.

Measured performance

JFK's “We choose to go to the Moon” speech: a 20-second excerpt and the 18-minute 15-second recording. Full recordings use 37 contiguous chunks, with every input sample retained. The M3 is a 16 GB MacBook Air.

The short Mac result and both phone results are three-run warmed medians. The full Mac row is one final-package verification pass; the prior three warmed passes with the same runtime/graphs had a 39.37-second median and identical tokens. Mac measurements exclude file I/O and preparation. Phone measurements include bounded file reading, chunking and output delivery, but exclude model preparation. The long phone runs ranged 51.20–57.71 s, with fair/serious thermal state after initial compilation; short phone runs were nominal. Measurements were made September 29–30, 2026.

Peak full-recording memory was 4.51 GB on Mac, combining simultaneous client and separately attributed ANE allocations. Phone client footprint peaked at 4.46 GB, which already includes its ANE allocation. These measurements do not reserve memory or guarantee that another app has the same memory allowance.

These are observed loads, not guaranteed clean-device compilation times. The Mac already had related audio/head graphs cached in the 346-second experiment. Packaging removes debug locations; its first Mac load took 60.14 s while reusing existing backend caches. Core AI's underlying caches are not fully observable, and the OS may invalidate them. The phone's four text assets took 151, 148, 113 and 116 seconds during its initial load.

Qualification

• Original checkpoint tensors were mapped exactly into the native Transformers reference, including the tied embedding/head values. Layer checks compared hidden states, KV updates and vocabulary decisions, including the compact 512-to-1,024-position cache transition.
• The final package returned identical tokens to the qualified runtime for every full-recording chunk. Phone full-recording outputs also matched Mac.
• Whisper-normalized WER was 61/2,220 errors (2.75%) on this recording, equal to the BF16 Transformers/MPS reference using the same cuts. The source took 204.60 s in one measured pass. This does not establish parity with native whole-recording recognition or a general accuracy ranking.
• A complete short phone trace matched 226 model-call scopes to ANE work, with zero app GPU intervals. Final-package Mac compiler metadata marks all seven assets fully on ANE. Full-length phone traces had incomplete signpost coverage, so they are not evidence of complete full-run placement. Placement evidence is not a measure of ALU utilization or energy efficiency.

Download the complete repository snapshot. Source paths, evaluation audio, reference tensors and compiled device caches are excluded. Hugging Face's snapshot metadata supplies file hashes for integrity verification.