VQV Signal
According to a post on X by @saltyph, a team built an inference engine for DeepSeek V4.1 Flash from scratch, using only Rust and CUDA with no PyTorch, SGLang or DSLs. The post reports about 1,740 tokens per second for decoding on 4x B300 and 2,060 tokens per second on 8x B300, plus a 1M-token prefill in about 12 seconds. The author says the code is on GitHub. The post had 61 likes and 38 bookmarks when captured.
According to a post on X by @saltyph, a team built an inference engine for DeepSeek V4.1 Flash from scratch, using only Rust and CUDA with no PyTorch, SGLang or DSLs. The post reports about 1,740 tokens per second for decoding on 4x B300 and 2,060 tokens per second on 8x B300, plus a 1M-token prefill in about 12 seconds. The author says the code is on GitHub. The post had 61 likes and 38 bookmarks when captured.
AI-assisted summary based on the listed source.
@saltyph said his team built an inference engine for DeepSeek V4.1 Flash from scratch, with a core of just Rust and CUDA and no PyTorch, SGLang or DSLs. They report about 1,740 tokens per second decode on 4x B300, 2,060 on 8x B300 and a 1M-token prefill in about 12 seconds, and the code is on GitHub; the post drew...
If the published code supports these claims, it would suggest that a small team can build a lean inference stack without the common Python frameworks. Builders and businesses evaluating serving options for large open models may want to review the GitHub repository and benchmark setup. The figures are self-reported in the post and have not been independently checked here, so reproduction results are worth watching.
VQV organizes public signals from inspectable sources. It does not independently verify the underlying report.
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
VQV surfaced this signal because it is recent, relevant to LLM Inference, connected to Scobleizer News Pulse (X).
No login, cookies, social SDKs, or automatic posting.
