TensorFold pushes Qwen to a million tokens on one GB10

TensorFold 1.2.1 opens a million-token context path for Qwen3.5 and Qwen3.6 on CUDA, while bringing MOSS-Image’s prompt encoder and image decoder into its native Metal pipeline. The November 1 release expands the community inference engine’s supported workflows, with specific hardware and backend restrictions.

The project reports that Qwen3.6-35B-A3B processed a 1,048,512-token prompt on one NVIDIA GB10 in 26 minutes and retrieved a passphrase placed 5% into it. That is a reported long-context retrieval check, not evidence that the model reasons reliably across every part of a million-token document. AITrending has not reproduced the run.

A larger window and new draft paths

The release says --rope-yarn 4, or a configuration containing a YaRN block, raises supported Qwen3.5 and Qwen3.6 CUDA windows to 1,048,576 tokens. It explicitly excludes HIP, Metal, DFlash2 drafters and other model families from that scaled-rope path. Memory still has to cover weights, context and runtime state.

Separately, Qwen3.6-35B-A3B gains a DFlash2 draft-model path. On a single GB10 with NVIDIA’s NVFP4 checkpoint and a 4.2K-token prompt, the project reports 130 tokens per second, compared with 99 using MTP. The release also adds Google’s MTP assistant for Gemma 4 31B on CUDA, reporting 28.3 tokens per second against 12.2 without drafts at one stream. These are project measurements for named configurations, not universal speedups or independent comparisons.

Native images on M5, a partial CUDA pipeline

For MOSS-Image, the Metal path now includes prompt encoding, denoising and final image decoding inside the native tools. The tagged recipe limits this implementation to M5 hardware with tensor units; it is not support for every Apple Silicon Mac. The documented case-building and reference-comparison workflow still uses Python tooling.

CUDA support has a different boundary. The native transformer was checked on RTX 5090 and GB10 hardware, while prompt encoding and image decoding remain in PyTorch. The image family does not register with TensorFold’s serving engine, so this release should not be read as a new image-generation HTTP endpoint.

Installing the engine

TensorFold is an Apache-2.0 Zig engine, separate from the models it loads. Its native language-model server reads supported checkpoints without requiring a Python or MLX runtime. The README documents Homebrew installation on Apple Silicon and release archives for Linux aarch64 and x86_64. The Linux CUDA archives require glibc 2.28 or newer and NVIDIA R580 or newer drivers for CUDA 13.0.

The server defaults to 127.0.0.1:8080, with API-key and restricted key-file options for authenticated access. Model weights retain their own licenses and download conditions. Version 1.2.1 also adds GET /v1/load under the metrics access rules, giving operators a view of queued work and backend capacity alongside the new inference paths.