What GPU do I need to run livekit/turn-detector?
A 135M-parameter speech model. 135M parameters, published in F32. View on Hugging Face
turn-detector is published by livekit on Hugging Face, with 696,553 downloads and 142 likes to date. It's a LlamaForCausalLM model built for text-classification, published natively in F32.
What turn-detector is
turn-detector is a 135M-parameter model published by LiveKit on Hugging Face that classifies audio/text signals, released under Custom license.
License note: a lab-specific license (tagged "other" on Hugging Face); read the model card's own license section before commercial use. Facts in this section are sourced from turn-detector's Hugging Face model card, not benchmarked by Aquanode.
What it's used for
- Speech synthesis
- Voice agent pipelines
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| FP32 | 0.5 GB | 0.6 GB | RTX 3070 | 1 | $0.088/hr |
| FP8 (quantized) | 0.1 GB | 0.2 GB | RTX 5060 Ti | 1 | $0.180/hr |
| INT4 (quantized) | 0.1 GB | 0.1 GB | RTX 3070 | 1 | $0.088/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run turn-detector at its published (F32) precision: 1× RTX 3070, at $0.088/hr per GPU ($0.088/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
turn-detector: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Grouped-query attention: every layer caches keys and values for a few shared KV heads. Cached per token: 23,040 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.09 GB |
| 8K (model maximum) | 0.18 GB |
Computed from the layer, head and window counts in the model's own configuration (livekit/turn-detector), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
turn-detector: common questions
How much VRAM does turn-detector need?
0.6 GB at FP32, 0.2 GB at FP8 (quantized), 0.1 GB at INT4 (quantized). All of those sit under 6 GB, the smallest capacity this page reasons about, so VRAM is not what limits where this model runs. The 0.5 GB of weights plus inference overhead is the whole requirement.
Can turn-detector run in 16-bit instead of FP32?
Yes. Its published weights are FP32, 0.5 GB, or 0.6 GB once inference overhead is added. Loading the same weights in 16 bits halves that to 0.3 GB, or 0.3 GB with overhead. How much accuracy the cast costs is model-specific and is not measured here.
How many copies of turn-detector fit on one RTX 3070?
13, by VRAM alone. That card carries 8.0 GB and one copy needs 0.6 GB at FP32, on a live rate of $0.088/hr for the whole card. Throughput is not modelled here, so 13 copies is not 13 times the requests served.
How to run turn-detector
Run turn-detector with Transformers (Python)
Generic example using Hugging Face's transformers library, not from the model's own docs.
from transformers import pipeline
clf = pipeline("text-classification", model="livekit/turn-detector", device="cuda")
print(clf("example input"))Deploy turn-detector on Aquanode
Aquanode has no one-click deploy template for turn-detector; you install the inference engine yourself with the commands below. Aquanode sells GPU pods billed per second, not a hosted inference API.
- Launch a bare GPU pod sized to the requirement above (1× RTX 3070 or larger).
- Open a terminal on the pod, or save one of the commands above as a startup script so it runs automatically the first time the pod boots.
- Run the command and connect to the resulting endpoint.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Qwen2.5 models
- Qwen2.5-0.5B-Instruct (494M, BF16)
- Qwen2.5-0.5B (494M, BF16)
- jina-code-embeddings-0.5b (494M, BF16)
- Qwen2.5-Coder-0.5B-Instruct (494M, BF16)
- Qwen2.5-Coder-0.5B (494M, BF16)
Lineage
Based on Qwen2.5-0.5B-Instruct.
Alternatives at this size
Other models for text-classification within about a third of turn-detector's 135M parameters, from other model lines.
- Julia-1 (144M, F32)
More on turn-detector
Fits on an 8 GB GPU at FP32: every model that fits in 8 GB.
Related reading: RTX 3070 pricing and specs, How much VRAM you need for LLMs, What AI inference is, and What AI model fine-tuning is.
Your own GPU cloud accounts
One console for the GPU clouds you already rent from. Your keys stay on your machine.
$20 a month, flat