Athena engine runs DeepSeek V4 + Qwen3.8 at 262K context on one DGX Spark

solyarisoftware · x · 2026-09-20

• A new inference engine called Athena has shipped for Nvidia GB10 systems, running DeepSeek V4 Flash and Qwen3.8 Flash-Next at 262K context on a single 128GB DGX Spark with 1,000 tps prefill.
• Benchmarks: DeepSeek V4 Flash hits 1,126 tps at 8K prefill and 948 tps at 256K (19.4 tps decode); Qwen3.8 Flash-Next does 1,071 tps prefill and 32.1 tps decode at 256K — prefill barely degrades from 8K to 256K.
• Key features: disk-cached long conversations (a 141,519-token session restores in 2.1s vs 2min20s fresh), speculative decoding, OpenAI/Anthropic-compatible API, tool calling, Qwen image/document input, persistent agent context, Docker install, and model switching without changing clients.

Original post →

More from Infra

• Agent Substrate Calculator Estimates Session Resume Speed Under Load — pjausovec · 2026-11-12
• Dell expands AI Data Platform: agents have a data-context problem, not a model problem — DavidLinthicum · 2026-11-12
• SGLang Megakernel Pushes Muse Spark to 1,109 tok/s, 2.1x Over Baseline — hsu_byron · 2026-11-12
• Potion UI's local login system turns your PC into a multi-user AI server — cocktailpeanut · 2026-11-12
• AWS Finally Built Its Own AWS Emulator — super_linear · 2026-11-12
• Megakernel is multiprogramming for GPUs: 'GPUs are entering the CPU age' — pranjalssh · 2026-11-12

Infra channel →