W1-JEV

Millisecond decisions. Less than 1¢ in GPU cost per 1,000 requests.

33 ms median latency · $0.00567 per 1,000 requests · 30 emails per second

中文模型介绍与评测报告

W1-JEV is a 4B-class language diffusion model for structured decision tasks, built on our W1 DLM. We applied post-training output-format alignment to reproduce JEV-style decision tasks, targeting decision workflows and high-volume automation.

On 231 matching public adaptation questions, W1-JEV answers 184 correctly, reaching 92% of the accuracy of the Jev 1.13.0 public record. Our reported cost comparison places GPU occupancy cost at approximately one seventh of the reference setup.

The measured run completed all 231 requests in 10.8373 seconds, with 100% strict output-format validity. At a GPU rate of $0.435/hour, the allocated cost is $0.00567 per 1,000 requests—about 0.57 US cents.

Built for high-volume decisions

When every input requires a decision, the latency and cost of each request add up. W1-JEV focuses on that repeated step to make batch decisions faster and more economical.

• Short waits. The reported run achieved 33.13 ms p50 end-to-end request latency and 21.32 requests/second throughput.
• Low GPU cost per decision. Measured GPU occupancy cost was $0.00567 per 1,000 requests at the stated hourly rate.
• Outputs that fit the workflow. All 231 completed requests passed strict format validation in this run.
• A local deployment path. The release includes a pure PyTorch model, a W1-specific runtime, a local HTTP service, and an executable choice-question client.

Accuracy: 92% of the reference accuracy

Results below cover 231 matching questions from the public adaptation set. W1-JEV results are publisher-reported; Jev 1.13.0 results come from the public record.

Metric	W1-JEV — publisher-reported	Jev 1.13.0 — public record
Correct answers	184 / 231	200 / 231
Raw accuracy	79.65%	86.58%
Simple / easy subset	48 / 48	48 / 48

W1-JEV matches the reference on the easy subset. Across the full matching set, it answers 16 fewer questions correctly, a 6.93 percentage-point accuracy gap.

“92%” is a relative accuracy ratio: (184 / 231) / (200 / 231) = 92%. W1-JEV's absolute accuracy is 79.65%; the ratio is specific to this set.

Performance and GPU cost

This run used a single NVIDIA GeForce RTX 5090, with 2 concurrent clients, one GPU worker and maximum batch size 2. The run comprised 116 batches (average batch 1.99), with up to 5 ms batching wait.

Metric	Reported result
End-to-end request latency, p50 / p95	33.13 / 374.02 ms
Throughput	21.32 requests/s
Wall-clock duration	10.8373 s
Requests completed	231 / 231
Strict format validity	100%
Peak PyTorch allocated VRAM	7.45 GiB
GPU hourly rate	$0.435 / GPU-hour
GPU occupancy cost per 1,000 requests	$0.00567

The cost is calculated from the reported run duration:

GPU occupancy cost per 1,000 requests
= (10.8373 / 3,600) × $0.435 × (1,000 / 231)
≈ $0.00567

This is the GPU occupancy cost allocated to the run, excluding other service overhead. It is not a hosted API price. Peak PyTorch allocated VRAM is an allocator measurement, not the total VRAM required for deployment.

Model architecture

W1-JEV uses a custom PyTorch LangDiT language diffusion transformer, with bidirectional attention, timestep conditioning, and adaptive layer normalization.

Property	Value
Model	Custom PyTorch LangDiT
Parameters	3,739,097,600 — approximately 3.74B, in the 4B class
Transformer blocks	48
Vocabulary size	64,512
Hidden / attention / FFN dimensions	2,048 / 3,072 / 7,168
Attention heads / head dimension	24 / 128
Attention	Bidirectional, RoPE, no causal mask
Conditioning	Timestep embedding and adaptive layer normalization
Default runtime input limit	4,096 tokens, configurable via model.max_seq_len
Default inference dtype	BF16

The checkpoint is stored in FP32; inference uses BF16 by default. The runtime input limit is configured through model.max_seq_len in config.json; RoPE positions are computed dynamically from the actual sequence length.

Runtime entry points

See Usage for download, installation, and a first request.

Use the repository's W1-specific runtime for inference. The main entry points are:

• structured_server.py — W1 runtime and local HTTP service.
• examples/ask.py — executable client for choice questions.
• config.json — model architecture and runtime context configuration.

Repository contents

File	Purpose
w1-jev.pt	FP32 model-only checkpoint; no optimizer, EMA, or other training state.
model.py	Pure PyTorch W1 language diffusion transformer.
config.json	Architecture and runtime context configuration.
tokenizer.json, tokenizer_config.json	W1 tokenizer and special-token metadata.
checkpoint.py	Strict validation of model-state keys, shapes, and finite values.
batch.py	Variable-length batching and answer-slot logits.
djev_template.py	Extracted djev schema, prompt, template, and scheduling helpers.
structured_server.py	W1-specific runtime and local HTTP service.
examples/ask.py	Executable choice-question client.
README.md	English model card.
report.zh.md	Chinese model card, benchmark report, and repository guide.