Zyphra ZAYA1-8B-Diffusion: 7.7x speedup (lossy) or 4.6x (lossless) via TiDAR

Zyphra released ZAYA1-8B-Diffusion-Preview, what it describes as the first MoE diffusion model converted from an autoregressive LLM rather than trained from scratch. The starting point is ZAYA1-8B, a MoE autoregressive model using Zyphra's CCA (Continuous Compression Attention) variant with CCGQA configuration. The conversion uses the TiDAR recipe across roughly 1.1 trillion additional tokens of mid-training: 600B tokens of diffusion-conversion training at 32k context, 500B tokens of native context extension to 128k, and a diffusion supervised fine-tuning phase. The headline speedup numbers are the news: a lossless sampler at 4.6x inference speedup with no systematic loss, and a logit-mixing sampler at 7.7x with some quality trade-off. Status is preview, not general availability — Zyphra describes the diffusion inference stack as "early-stage."

The mechanism is single-step speculative diffusion with order-constrained generation: instead of full random-position masked diffusion, the model generates contiguous subsequences extending from the prefix, predicting 16 tokens simultaneously per forward pass with a shared KV-cache across the token block. That shifts decoding from memory-bandwidth bound to compute-bound — which matters because modern accelerators have been scaling FLOPs faster than HBM bandwidth for several generations, and inference is increasingly bottlenecked on memory, not arithmetic. On AMD MI300x they report roughly 3 block proposals per pass; on the newer MI355x roughly 5. The order-constrained framing also means this is not a free-form diffusion model in the image-gen sense — it's closer to large-block speculative decoding with a diffusion-style training objective than to "diffusion language model" in the strongest sense of that phrase.

The honest evaluation reading is two-layered. First, Zyphra emphasizes "pass@" metrics rather than standard accuracy benchmarks because this is a base mid-train checkpoint pre-RL training; gains are reported on LCB-v6 with "minimal evaluation degradation" versus the autoregressive base, but no per-benchmark delta tables appear in the announcement. Second, the dual-sampler reporting — 4.6x lossless and 7.7x with trade-off — is the right shape of disclosure, but the size of the trade-off at 7.7x is not quantified in the public release. Builders evaluating this should read both numbers: the lossless figure is the conservative claim, the headline 7.7x is the aggressive claim, and the actual decision on whether to use the logit-mixing sampler depends on tolerance for quality variance on your workload. ZAYA1-8B-base (the autoregressive model) is on Hugging Face; the diffusion variant's release artifacts and license status are not fully detailed in the announcement.

For builders watching inference economics: if the 4.6x lossless number holds up in third-party benchmarking on real workloads at modest batch sizes, this is a meaningful change in the cost curve for high-volume text generation, especially on AMD silicon where the MI300x/MI355x numbers were measured. The architectural claim — converting an AR model rather than retraining from scratch — is also methodologically interesting because it suggests existing AR MoE checkpoints could be retrofitted into diffusion variants without re-running expensive pretraining, if the TiDAR recipe generalizes outside Zyphra's stack. The tests that will decide whether this is a permanent change or a single-vendor research preview are reproductions on other AR MoE bases (Qwen MoE, DeepSeek MoE variants), and clean per-benchmark numbers on standard evaluations once Zyphra moves past the pre-RL checkpoint.