SGLang adds support for gpt-oss-120b on Hopper and Blackwell GPUs

When OpenAI releases an open model, engineers' first question is not about the quality of its answers. It is something else: what hardware will run it, and how much will a million tokens cost? Until a model has proper support in an inference engine, it remains an attractive line in a press release. That is why the news from LMSYS Org may look modest but matters to practitioners: SGLang now supports gpt-oss-120b on Nvidia's Hopper and Blackwell GPUs.

What was announced

The LMSYS team, which develops SGLang, announced support for gpt-oss-120b on two architectures at once: Hopper (H100, H200) and Blackwell (the B200 family). Eigen AI Labs, Nvidia, the SGLang developers themselves and the open-source community contributed to the work. The post also says support arrived within a few days of the model's release, but the exact number is cut off in the available text, so we are not giving a figure.

Here is the context. gpt-oss-120b uses a Mixture of Experts architecture: only a small portion of its parameters is activated for each token. According to OpenAI, it has roughly 117 billion parameters in total, with around 5 billion active. The weights are released under the Apache 2.0 license and quantized in MXFP4 format, allowing the model to fit on a single 80 GB GPU. Models like this are particularly sensitive to the quality of the engine: both speed and cost depend on how it distributes experts and uses low-precision kernels.

Why this matters in practice

Imagine a typical situation. You are building an internal developer assistant or an agent that navigates documentation and calls tools. You would rather not send code and internal data to an external API. You have your own GPUs or rent them from a cloud provider. Next, you need to choose how to serve the model: vLLM, TensorRT-LLM, SGLang or something else.

Support in SGLang adds another viable option to that list. SGLang is known for efficiently reusing prompt prefixes, which is particularly helpful for agent workflows where every request repeats a long system prompt and conversation history. For gpt-oss-120b, positioned as a model for reasoning and tool use, this is a typical workload.

Blackwell deserves a separate mention. Software support often lags behind new hardware, leaving owners of new B200 GPUs waiting months for optimized kernels. When support is announced for both Hopper and Blackwell from the outset, teams do not need to maintain two separate stacks for older and newer cards.

What the post does not include

The limitations need to be stated clearly. The announcement contains no figures: no throughput in tokens per second, no latency measurements and no comparisons with other engines. There is also no data on how the model performs with long contexts and high concurrent workloads. This is a support announcement, not a benchmark, and it cannot serve as the sole basis for a migration decision.

Also, "supported" does not mean "optimal." Early versions of an integration usually work correctly but do not always get the most out of the hardware. Performance improvements arrive in subsequent releases, so it makes sense to follow SGLang updates rather than pin the first version forever.

What to do

If you already host LLMs yourself, run a small experiment. Take a real request stream or a recording of one, deploy gpt-oss-120b in SGLang and in the engine you currently use, and compare three things: time to first token, tokens per second under load, and hourly GPU costs per thousand requests. Test tool calling and structured output separately, because those are where differences between engines most often surface.

Even if you do not own hardware, the news is still useful: the more engines can serve an open model, the stronger the competition among inference providers and the lower the price per token. Choosing open weights becomes less of a gamble when a model has several mature ways to run it.