GLM-ASR-Nano-2512

GLM-ASR-Nano-2512 is a compact, open-source automatic speech recognition (ASR) model from Z.ai, weighing in at 1.5 billion parameters. It belongs to the GLM family and targets a specific pain point: accurate transcription in real-world conditions that trip up larger, more general-purpose models. The model is a dense architecture—no mixture-of-experts tricks—meaning all 1.5B parameters are active during inference. That makes it straightforward to run on consumer hardware without the complexity of managing expert routing.

What sets this model apart is its focus on Chinese dialects (particularly Cantonese) and its ability to handle whisper-quiet speech—audio so low-volume that most ASR systems return silence or garbage. On standard benchmarks like Wenet Meeting (noisy, overlapping meeting speech) and Aishell-1 (Mandarin), it achieves the lowest average error rate (4.10) among comparable open-source models, outperforming OpenAI’s Whisper V3 despite being significantly smaller. Licensed under MIT, it’s free to use, modify, and deploy.

GLM-ASR-Nano-2512 uses a dense encoder-decoder architecture with 1.5B total parameters. Unlike MoE models where only a subset of parameters activate per token, this model loads and runs all parameters at inference time. The tradeoff is straightforward: you get predictable memory usage and consistent latency, but you pay the full VRAM cost for every forward pass.

Because Z.ai has not published a context length, we treat it as the default for the transformers integration—typically 2048 or 4096 tokens for audio inputs. In practice, the model processes audio frames directly through a processor that handles resampling and feature extraction, so the effective context is dictated by audio duration rather than text tokens. For most real-world use cases (meetings, interviews, short recordings), this is more than sufficient.

The model is designed to be run with Hugging Face transformers (from source, as of early 2025) and will soon support vLLM and SGLang for optimized inference. The processor includes a apply_transcription_request method that wraps audio preprocessing and chat template formatting, making integration trivial.

GLM-ASR-Nano-2512 excels in three specific areas:

• Cantonese and Chinese dialect recognition. Standard Mandarin models often fail on Cantonese due to tonal and phonetic differences. This model was explicitly trained to bridge that gap. If your workflow involves transcribing Cantonese interviews, customer calls, or media content, this is one of the few open-source options that delivers usable accuracy.
• Whisper/quiet speech. The model was trained on low-volume audio where speakers are barely audible—think whispered conversations, far-field microphones, or soft-spoken users. Traditional ASR models (including Whisper V3) tend to miss or hallucinate in these conditions. GLM-ASR-Nano-2512 captures and transcribes them reliably.
• Real-world meeting transcription. The Wenet Meeting benchmark includes overlapping speech, background noise, and varied acoustics. The model’s average error rate of 4.10 on that benchmark is the lowest among open-source models at any size, making it a strong candidate for on-premise meeting transcription systems.

Beyond these strengths, it supports 17 languages with WER ≤20%, including English, Japanese, Korean, and several European languages. However, its primary advantage is in Chinese-language tasks. Use it for voice assistants, dictation software, accessibility tools, and any application where local, low-latency ASR is required and cloud APIs are not an option.

This is where the model shines for practitioners. At 1.5B parameters, it fits comfortably on consumer GPUs without quantization. Here are the practical numbers:

• Minimum VRAM (FP16): ~3 GB for model weights alone. Add ~1 GB for processor overhead and audio buffers, so 4 GB total. An RTX 3060 12GB or M1/M2 Mac with 8GB unified memory can run it.
• Recommended VRAM: 6–8 GB for comfortable inference with batch size 1 and no quantization. An RTX 4060, RTX 3070, or M4 Max (16GB) will handle it easily.
• Quantization: For most users, Q4_K_M via llama.cpp or Ollama is the sweet spot. It cuts VRAM to ~2 GB and runs on 4 GB cards (RTX 3050, M1 base). Q8_0 offers near-lossless quality at ~3.5 GB VRAM. Q2_K is possible but not recommended—error rates increase noticeably.
• Expected performance: On an RTX 4090 (FP16), expect 80–120 tokens per second for audio transcription (depending on audio length). On an M4 Max (FP16), roughly 50–70 tokens/second. Quantized (Q4_K_M) on an RTX 3060, you’ll get 30–50 tokens/second—more than enough for real-time transcription.

Quickest way to start: Ollama. Pull the model (once available in the Ollama library) or convert the Hugging Face checkpoint to GGUF format and import it. Otherwise, use transformers from source as shown in the model card:

1from transformers import AutoModelForSeq2SeqLM, AutoProcessor
2
3processor = AutoProcessor.from_pretrained("zai-org/GLM-ASR-Nano-2512")
4model = AutoModelForSeq2SeqLM.from_pretrained("zai-org/GLM-ASR-Nano-2512", dtype="auto", device_map="auto")
5
6inputs = processor.apply_transcription_request("audio.mp3")
7outputs = model.generate(**inputs, do_sample=False, max_new_tokens=500)
8print(processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True))

No cloud dependency. No API keys. Just local inference.

The most direct competitor is OpenAI Whisper V3 (1.5B variant). On standard English benchmarks, Whisper V3 holds a slight edge. But on Chinese dialects, quiet speech, and noisy meetings, GLM-ASR-Nano-2512 consistently wins. Whisper V3 also has a larger context window (224k tokens for audio) and supports 99 languages, but its performance on low-volume audio is mediocre. If you need broad multilingual coverage and can tolerate higher error rates on edge cases, Whisper V3 is the safer bet. If your primary language is Chinese (especially Cantonese) or you deal with whispered input, GLM-ASR-Nano-2512 is the better choice.

Another alternative is SenseVoice (by Alibaba), which also targets Chinese ASR but is larger (around 1.8B parameters) and less optimized for quiet speech. GLM-ASR-Nano-2512 matches or beats it on benchmarks while being easier to run locally due to its smaller size.

For English-only workloads with clean audio, neither model is ideal—use a dedicated English ASR model like Whisper small or Wav2Vec2. But for the niche this model fills, it’s the most capable open-source option available today.