A Reddit user ran local tests with the same quantized Qwen3.8 FlashNext model: Basalt delivered 110–170 tokens per second, while the previously used Strata reached 70–110 tokens per second. What matters here is not merely a change of parameters, but the shift in local LLM inference tools toward a more stable engineering experience; however, the results come from a single user and cannot yet be treated as a universal conclusion.

What this is

Basalt and Strata are both local inference engines for running mixture-of-experts models (MoE, which assigns different tasks to subsets of parameters). They enable models to generate text on a personal computer or server, with output speed typically measured in tokens per second.

After moving from Strata to Basalt, the poster said Basalt was faster and less prone to the model forgetting the current task or drifting in its reasoning after context compression. Context compression means condensing older information into a summary when a conversation becomes too long, thereby saving memory.

Industry view

We believe this feedback signals a clear transition: competition among local models is moving from “whether they can run” to “whether they are useful to use.” Speed, stability, and compatibility with hardware and quantized models may influence developer choices more directly than model leaderboard scores.

The risks are equally clear. The test did not disclose its complete environment, and no third party has reproduced the results. Moreover, a single GPU paired with more than 100GB of VRAM, memory, and storage is far beyond the configuration of most ordinary consumers. Better tools can narrow the experience gap, but they cannot yet eliminate deployment costs.

Impact on regular people

For enterprise IT: As local inference engines become more mature, they can help enterprises evaluate model deployments that keep data entirely inside the network. At this stage, however, dedicated hardware, model optimization, and operational expertise remain necessary.

For individual professionals: Higher speeds should make local processing of long documents, code, and sensitive materials smoother. Whether users can adopt these tools still depends on their hardware and whether they are willing to take on deployment and maintenance.

For the consumer market: Improved runtime quality could drive the emergence of more local AI applications, but the high barrier to entry means that, in the near term, these tools are more likely to reach professional workstations than become a default capability in mainstream software.