Alibaba says Qwen4-Max activates 71 billion parameters per token against a 2.4 trillion-parameter backbone, with a separate short-sequence embedding system. Three recurrent layers compress earlier context for each sparse-attention layer. On the company’s Work360 test, average task cost fell from $4.43 to $1.67 against its prior Max model. Both token usage and price changed. The advertised $0.075 per million input tokens applies to implicit-cache hits, not the full workload.

Zyphra’s open-weight Zonos 3 Mini takes 15 sequential audio steps to produce a second of speech at 44.1 kHz. Its codec represents each frame with one semantic codebook and nine for acoustic detail; a decoder supplies the waveform. Zyphra reports under 570 ms to first audio in a Pixel 9 browser and under 150 ms on an M3 MacBook Air. Peak aggregate server throughput is not the number a listener waits for.

Sources: qwen.ai · alibabacloud.com · huggingface.co