llm.buzz is part of Surogate.

• SurvivesA crashed worker replays the session's log and carries on.
• ContainedCode runs in the agent's own sandbox pod.
• GovernedEvery tool call passes a policy frozen per session.
• Multi-tenantStorage, keys, memory and limits stay with the tenant.

Training throughput against Unsloth

2.53×

Surogate 53,900 tok/sUnsloth 21,300 tok/s

1× H100 80 GB. Qwen3-0.6B LoRA, BF16 on both: 53,900 against 21,300 tokens/s.

Serving one user

802tok/s

Surogate 802 tok/svLLM 346 tok/s

1× RTX 5090. Qwen3.5-0.8B, GGUF Q4_K_M, decode. vLLM (NVFP4): 346 tok/s.

Training on four cards

136,200tok/s

Four cards 136,200 tok/sOne card 36,400 tok/s

4× RTX 5090. Qwen3-0.6B, FP4 LoRA. 36,400 tok/s on one card.

llama.cpp's aggregate decode

7.0×

Surogate 292.9 tok/sllama.cpp 41.9 tok/s

8× RTX 5090. GLM-5.3-Flash GGUF, a 200 GB MoE, 16 users: 292.9 against 41.9 tok/s.

Selected results from the engine's training and serving benchmarks, as its README publishes them. RTX 5090 serving rows were measured from 2026-08-30 to 2026-09-07 with the cards at a 400 W limit, against vLLM 0.27.1 and llama.cpp CUDA builds. The README also lists the rows where the other engine leads.

• Surogatesurogate.ai
• Surogate Studioops.surogate.ai
• Documentationdocs.surogate.ai
• The enginegithub.com/invergent-ai/surogate
• Surogatesgithub.com/invergent-ai/surogates