What happened
Tencent’s GitHub commit for Hy-Nomas is titled “inference: 1) add B300 support; 2) SP perf optim; 3) add TP support.” It changes inference code, distributed utilities, launch scripts, tests, application code, and documentation. The update is a repository-level implementation change rather than a hosted-product release, pricing announcement, or change in model access.
The README’s standard inference example now passes --sp-size 1 --tp-size 8. Tencent’s documentation describes this preset as tensor-sharding dense DiT attention and MLP projections over eight ranks. It directs operators to a tensor-parallel inference guide for topology overrides, tests, and current limitations. This makes tensor parallelism an exposed deployment control in the documented path, rather than an implementation detail hidden behind a fixed launch arrangement.
Tencent also wires those controls into the Gradio application. Its command-line interface accepts --sp-size and --tp-size, transfers them to the parallelism configuration, and validates them against the distributed world size. For a Gradio deployment, the code raises an error unless WORLD_SIZE equals sp_size * tp_size. When more than one process is used, the application initializes a world mesh; it also records both settings in inference metadata.
Several launch examples move from sequence parallelism alone to a combination of sequence and tensor parallelism. The application documentation uses two hosts with eight GPUs each as an example, configured with --sp-size 2 --tp-size 8. It says that sp_size × tp_size should equal the host count multiplied by NPROC_PER_NODE. The commit adds a TP_SIZE environment-variable override and distributed tensor-parallel tests, as well as B300-specific inference code and a B300 input utility.
The README continues to identify a low-memory inference configuration for a single NVIDIA GeForce RTX 4090. That path and the new B300 work describe two different deployment targets: a single-card configuration for lower-memory use and distributed configuration for multi-GPU execution. Tencent’s primary announcement is the Hy-Nomas commit d91fc49.
What the archive adds
New Runtime covered Tencent’s official SGLang inference instructions for Hy-Nomas on 2026-11-01. This is a new event in the same project: the earlier note concerned a documented SGLang usage path, while this commit adds B300 targeting and tensor-parallel inference configuration.
Trend signal
This update weakly supports agent economics moves to completed work. Distributing generation inference across GPUs is an engineering route to making workloads operationally feasible. But Tencent’s commit does not state cost per completed task, fleet-level throughput, or a comparison of completed-generation economics, so it is not evidence of a shift in the metric teams should optimize.
Connections to other events
An earlier New Runtime note from 2026-11-06 covered Hyperstition’s Hyperkernel release for NVIDIA B300. Both events target the same hardware generation, but at different layers: Hyperkernel concerns GPU kernels, while Tencent’s change exposes B300 support and parallelism controls in a video-model inference stack.
Why it matters
For teams operating Hy-Nomas, the immediate value is a clearer topology contract. Operators can configure sequence and tensor parallelism together instead of treating sequence parallelism as the sole visible setting in the launch examples. The documented two-host example and the WORLD_SIZE validation give a concrete starting point for composing a process group, reducing the chance of a launch whose declared parallel dimensions do not match its actual worker count.
The single-RTX 4090 low-memory configuration also gives developers a named option for local evaluation before committing a multi-GPU environment. Neither that configuration nor B300 support establishes that a particular prompt, video size, latency target, or output-quality requirement will work in a given deployment. The meaningful change is controllability: Hy-Nomas now makes the distribution dimensions explicit enough to test and tune.
Limits and uncertainty
Tencent’s commit documents support and configuration, but it does not establish performance across different prompts, video sizes, hardware fleets, or operating conditions. The SP=1, TP=8 preset is a stated configuration, not a general speed or cost guarantee. Teams still need to consult the linked tensor-parallel guidance, validate their network and process topology, and run their own output-quality and latency tests.
The commit also does not state serving prices, B300 availability, comparative throughput, or fleet-level scaling results. Compatibility under alternative sequence- and tensor-parallel combinations remains a deployment question. The RTX 4090 low-memory path identifies a configuration to evaluate, not a promise that every Hy-Nomas workload fits or performs adequately on that card.
Skill to build now
Turn parallelism settings into an explicit deployment check. Before launching a distributed Hy-Nomas job, calculate sp_size * tp_size from the intended topology and compare it with the actual WORLD_SIZE; use the two-host, eight-GPU-per-host example as a template only when its 16-process shape matches the cluster. Then run a fixed prompt set under at least two configurations and capture generation time, failures, GPU memory use, and output acceptance criteria. This separates a valid topology from a useful production configuration.
Reader takeaway
Hy-Nomas now exposes tensor parallelism as a documented inference control, with an SP=1, TP=8 preset and B300-oriented code. Teams using the project should treat it as a concrete deployment option: match the process world to the parallelism dimensions, then measure the result on their own generation workload rather than assuming a universal performance gain.