curated · 2026-08-07

This is worth a click because these aren't lab benchmarks — they're the actual kernels Tencent Hunyuan runs in production, now merged straight into SGLang's main branch. The three optimizations target real MoE serving pain points: the Attention kernel uses dynamic scheduling to stop GPU idle time when decode requests have wildly different context lengths, averaging 2.25× faster than FlashInfer/FlashAttention. The Router GEMM runs in BF16, hitting 1.3–3.22× over FP32 cuBLAS with much lower error than TF32 cuBLAS. The MoE kernel fuses small expert matrix multiplies, reaching 4.21× over Triton on Qwen3. End-to-end, Hy3-FP8 on 8×H20 saw TPOT drop 15–48.8%. The caveat I'd flag: all numbers are on H20 and H200, and the end-to-end tests use Tencent's own Hy3 and LongCat models. If you're serving a different MoE model, you'll need to benchmark yourself. But the code is already in SGLang — install and go, the barrier is low.

The headline feature is 50 character references in one generation. That lets you build crowd scenes or multi-character setups without stitching clips together. The 30-second cap plus built-in audio and dialogue is practical for short ads or storyboards—it skips the separate dubbing step. But Runway only posted a tweet. No word on how long a 30-second clip takes to render or what it costs. Their pricing has never been cheap, and if 50 characters multiplies the bill, this feels aimed at teams with real budgets. I'd wait for someone to post speed tests and invoices before getting excited.

The reason this post matters: when Fable 5 launched, Anthropic blocked nearly all biology queries to prevent misuse—even routine lab result checks or symptom questions got punted to the dumber Opus 5. They've now brought in internal and external experts to rewrite the classifier's constitution, carefully carving out benign use cases and retraining on the new rules. Result: about 85% fewer biology-related fallbacks. Don't read this as "Fable 5 is now open for biology research." Dual-use topics—virology, toxicology, molecular design—remain fully blocked. Professional research and drug development are still off-limits. Anthropic says they're working on trusted access pathways to close that gap, but no timeline yet. I'd call this a practical classifier tune-up, not a safety pivot. Everyday users and clinicians get less friction, but the gate on frontier bio capabilities stays shut.