Quantized MLX weights for beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - model.safetensors - config.json - tokenizer.model - tokenizerconfig.json - preprocessorconfig.json - specialtokensmap.json - keymap.json - conversionsummary.json This checkpoint has been re-validated against the current Swift and Python MLX runtimes. Verified semantic parity on an English fixture: - official CUDA reference path (transformers native Cohere ASR) Fastest and smallest, but introduces a lexical regression on the repo sample (Kaldi → Khaldi). - Generated from the Swift-compatible fp16 checkpoint beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - This repository contains inference artifacts only. Refer to…
Open-weight model · Speech recognition
confucius4-r2t2-bf16-mlx
by App Automaton appautomaton/confucius4-r2t2-bf16-mlx
confucius4-r2t2-bf16-mlx is an open-weight model for speech recognition from App Automaton, released under other. It has 2B parameters. At 16-bit it needs about 4.9 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
MLX-native bf16 conversion of NetEase Youdao's Confucius4-R2T2, a streaming speech recognition model finetuned from Qwen3-ASR-1.7B. It runs through the mlx-speech runtime on Apple Silicon with no PyTorch, vLLM, or cloud API at inference time.
Parameters2B
Context—
Weights4.1 GB
Licenseother
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve confucius4-r2t2-bf16-mlx (2B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 4.1 GB | 4.9 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 2.0 GB | 2.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 1.0 GB | 1.2 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
confucius4-r2t2-bf16-mlx on every accelerator the SAVRN Index prices, at every precision
Model Card
MLX-native bf16 conversion of NetEase Youdao's Confucius4-R2T2, a streaming speech recognition model finetuned from Qwen3-ASR-1.7B. It runs through the mlx-speech runtime on Apple Silicon with no PyTorch, vLLM, or cloud API at inference time. Streaming — feed microphone PCM as it arrives: Or download once and load by path: mlx-speech implements NetEase's chunk loop: prefix rollback, per-window token budget, and append-only commit. Each window reuses the previous window's mel frames, closed audio-encoder blocks, and decoder KV prefix, and recomputes only what changed. - This repo contains the MLX runtime artifact only (no PyTorch checkpoint). The model weights are licensed under the NetEase…
Excerpt from the card by App Automaton, licensed other.
Configuration
- Architecture
- Qwen3ASRForConditionalGeneration
- Model type
- qwen3_asr
Identity and Version
- Repository
- appautomaton/confucius4-r2t2-bf16-mlx
- Publisher
- App Automaton
- Task
- Speech recognition
- Modality
- Audio
- Library
- mlx
- Parameters
- 2B parameters
- Languages
- zh, en, yue, ar, de, fr, es, pt
- Revision
- 00c86165790fe613f20904cb60a6698ee9968d1e
- First published
- 2026-09-27
- Last updated
- 2026-09-27
Files and Weights
14 files, 4.1 GB in total. The weights are 1 file totalling 4.1 GB in safetensors.
Weights1 file · 4.1 GB
Configuration6 files · 10.0 KB
Tokenizer4 files · 15.9 MB
Documentation2 files · 14.7 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 4.1 GB | 6234ed2e1e83 |
| added_tokens.json | Configuration | 1.6 KB | — |
| chat_template.json | Configuration | 1.2 KB | — |
| config.json | Configuration | 5.8 KB | — |
| generation_config.json | Configuration | 142 B | — |
| preprocessor_config.json | Configuration | 330 B | — |
| special_tokens_map.json | Configuration | 1.0 KB | — |
| LICENSE | Documentation | 11.1 KB | — |
| README.md | Documentation | 3.6 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| merges.txt | Tokenizer | 1.7 MB | — |
| tokenizer.json | Tokenizer | 11.4 MB | 049960271416 |
| tokenizer_config.json | Tokenizer | 12.5 KB | — |
| vocab.json | Tokenizer | 2.8 MB | — |
License and Download
- License
- other
- Access
- Open weights, no gate
- Download size
- 4.1 GB
Download from App Automaton
Released by App Automaton through its official repository on Hugging Face.
Built From
- Derived from netease-youdao/Confucius4-R2T2
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 4.1 GB |
| 16-bit | 4.1 GB |
| 8-bit | 2.0 GB |
| 4-bit | 1.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About confucius4-r2t2-bf16-mlx
How much GPU memory does confucius4-r2t2-bf16-mlx need?
About 4.9 GB at 16-bit and 1.2 GB at 4-bit: the weights (2B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run confucius4-r2t2-bf16-mlx on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What license is confucius4-r2t2-bf16-mlx released under?
other, as its publisher declares it. Read the license text before commercial use.
Similar Models
Quantized MLX weights for beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - model.safetensors - config.json - tokenizer.model - tokenizerconfig.json - preprocessorconfig.json - specialtokensmap.json - keymap.json - conversionsummary.json This checkpoint has been re-validated against the current Swift and Python MLX runtimes. Verified semantic parity on an English fixture: - official CUDA reference path (transformers native Cohere ASR) Matches fp16 on the repo sample while reducing memory substantially. - Generated from the Swift-compatible fp16 checkpoint beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - This repository contains inference artifacts only. Refer to the upstream Cohere model…
SeamlessM4T is our foundational all-in-one Massively Multilingual and Multimodal Machine Translation model delivering high-quality translation for speech and text in nearly 100 languages. SeamlessM4T models support the tasks of: - Automatic speech recognition (ASR). - 101 languages for speech input. - 96 Languages for text input/output. - 35 languages for speech output. We are releasing SeamlessM4T v2, an updated version with our novel UnitY2 architecture. This new model improves over SeamlessM4T v1 in quality as well as inference speed in speech generation tasks. The v2 version of SeamlessM4T is a multitask adaptation of our novel UnitY2 architecture. Unity2 with its hierarchical…
Pure-MLX selective-int8 conversion of IBM's for local automatic speech recognition on Apple Silicon. The complete waveform-to-text runtime uses mlx-speech; PyTorch, Transformers, mlx-audio, vLLM, and cloud APIs are not runtime dependencies. Portuguese, and Japanese The Granite causal language model uses 8-bit affine weight quantization with group size 64. The acoustic encoder, QFormer projector, normalization, convolution, biases, activations, and KV cache retain BF16 precision. The exact quantization contract is stored in config.json, and the runtime reconstructs the saved quantized module set from checkpoint.scales tensors before strict weight loading. BF16/int8 checks used two 12–13…
The Qwen3-ASR family includes Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, which support language identification and ASR for 52 languages and dialects. Both leverage large-scale speech training data and the strong audio understanding capability of their foundation model, Qwen3-Omni. Experiments show that the 1.7B version achieves state-of-the-art performance among open-source ASR models and is competitive with the strongest proprietary commercial APIs. Here are the main features: Novel and strong forced alignment Solution: We introduce Qwen3-ForcedAligner-0.6B, which supports timestamp prediction for arbitrary units within up to 5 minutes of speech in 11 languages. Evaluations show its timestamp…
MLX-native bf16 conversion of Qwen3-ASR-1.7B for local automatic speech recognition on Apple Silicon — English, Chinese, and mixed Chinese/English. It runs through the mlx-speech runtime with no PyTorch and no cloud API at inference time. Weights ship as plain.safetensors. Install mlx-speech, then load by alias (downloads on first use): Or download the weights once and load by local path: - This repo contains the MLX runtime artifact only (no PyTorch checkpoint). - v0 is single-pass offline ASR; streaming, timestamps, and long-audio chunking are deferred. - For a smaller download, use the int8 build. Apache 2.0, following the upstream Qwen3-ASR release. Refer to the original Qwen release…