Frank

TensorFold 1.1.4 is live. Up to 5× faster Qwen3.8-27B cold prefill on GB10, native Gemma 4 on CUDA, faster Flash Next multi-stream serving, and a long list of fixes from the community. What's new and fixed: • Qwen3.8-27B on CUDA has a new default prompt/decode recipe. On GB10, cold prefill improves by 54% at 2k, 107% at 8k, 258% at 32k and 416% at 64k. The 64k result goes from 203 to 1,045 tok/s; decode gains about 2% across 1 to 8 streams. Optional int8 prompt precision is also available, with bf16 remaining the default. • Gemma 4 12B and 31B now run on native CUDA from MLX affine checkpoints, with their drafters. On GB10, the 12B's MTP assistant takes single-stream greedy chat from 30 to 67 tok/s, and z-lab's DFlash drafter takes the 31B from 12.6 to 29.9 tok/s. Image paths are implemented, but their GB10 image test cells have not run yet. • Flash Next on CUDA verifies multiple streams' drafted windows in one pass. The GB10 served sweep gains 35%, 75% and 116% at 2, 4 and 8 streams, with tokens unchanged. Greedy MTP heads batch together; sampled heads, trees and media retain their per-stream path. • More kernel work: fused Flash Next DeltaNet prefill, a reused PLE worker pool, Qwen attention reads that reuse KV data in L2, segmented split-K decode, and fewer Nemotron RADIO launches. GLM-5.3's Metal sparse-index scan drops from 1.59 ms to 0.14 ms in the measured cell. Startup timing tables also stop one noisy measurement from setting later performance. • Drama 3 CUDA prompt chunking is opt-in. With 128-row chunks on GB10, first audio drops from 18.2 s to 0.76 s and the full utterance from 47.5 s to 30.2 s. The default remains unchanged; larger chunks can change rounding. • Opt-in speech controls now include silence-tail, onset and mid-utterance guards; whole-reply redraws with budgets and deadlines; short-reply and register guards; a controls allowlist; synthesis frame limits; and attested voice catalogs. The voice tool can build those catalogs from reference audio and transcripts. Drama CUDA replay also waits for buffer fills before reuse. • Nemotron-H MLX 4-bit now runs on the opt-in native HIP backend, including MTP, prompt caching and tensor parallelism. Unsupported flags are refused before loading, and Qwen HIP can cancel during long prompt chunks. • Idle unload can free a model's memory after a configured idle period and reload it on the next request. A failed reload returns a retryable error; ranked GLM CUDA engines refuse this option. • Sticky CUDA faults now make /health return 503. Driver-refused CUDA graphs can run eagerly instead of spinning, and queued writes retain their destination buffers through allocation failures. • GLM's draw ring no longer fills after 64 completed or cancelled cold requests. Long chats retain less tokenizer/template scratch, kept-alive connections retain less memory, and a missing disk state no longer evicts cached states before its read fails. • Memory and cancellation fixes cover Drama Metal loads that exceed the budget, Nemotron Metal Mamba-pool sizing and released slots after refused work, Qwen draft-node counts above 15, and prompt passes that now stop between chunks on cancellation or shutdown. GLM on one Mac also keeps its draft-acceptance estimates steadier after missed drafts. • Flash Next CUDA gains POST /v1/decisions on one GPU and loads valid nested checkpoint shards. Qwen CUDA vision reads Hugging Face-named towers. Managed benchmarks can select bf16 or fp8 KV cache and record the choice. Windows builds with nvcc and cl.exe work again. • Slide learning counts only kept lessons. App fixes cover engine-refused checkpoints, TUN/fake-IP proxies, Google's sign-in callback and visible setup errors. The docs now have a native GLM-5.3-Flash CUDA page. • Experimental DeepSeek work adds native prompt rendering, HTTP 400 responses for template refusals and four-rank prefill planning. It remains behind the whole-model gate; release binaries do not serve DeepSeek checkpoints yet. Release checks cover cold prompts from 2k to 64k, 1/2/4/8-stream sweeps and reference checks, with Linux and macOS host tests passing. The Qwen recipe changes prompt arithmetic, so output bits can differ from 1.1.3; its quality check matches the previous recipe at 27/40 exact cases, and concurrent/resumed checks still hold. One measured regression is documented: Nemotron Lightning greedy chat at four streams on GB10 is 3.8% slower. All three release archives are signed and verified against the site's trust root. The nine published assets were downloaded back and checked against the signed files. Homebrew is updated and upgrade-tested, and a fresh install through the website script verified the archive signature. A huge thank you to the release contributors: @bobclawblaw @plotarmordev @rielstamand @MiaAI_lab @Bizuayeu @jishvenugopal @petruspennanen @wesley4546 And GitHub contributors jschmied, Mirrdhyn, wojo, heitke, BlivionIaG, harrisonfriia and paragontasx. Thank you for the reports and requests, GitHub users Chris9121, flashosophy, Gogo6969, Idoo689, johncarterlives, mikog8tor, samwang0041-star and tonghuaxuan2016-dot. Attribution also matters. The Gemma family builds on earlier work by gilby on GitHub and @codengod. Credit to Google for the Gemma assistant, z-lab for DFlash and the 31B drafter, MLX/mlx-lm and Hugging Face Transformers for model references, and Fish Audio, SGLang-Omni and Descript for the Drama model/codec references. Full source and licence credits are preserved in THIRD_PARTY_NOTICES.md and docs/credits.md. And thank you to everyone testing, sharing measurements, contributing fixes and giving me access to hardware. This project would not be moving this quickly without you. @MiaAI_lab @plotarmordev @petruspennanen @volatilemarkts tensorfold.dev