Community
Developer runs eight 250K-context coding agents on two RTX 3090s with RAM-tiered sparse attention
A Reddit user reports serving Qwen3.8-27B W4A16 to eight agents at 250K context each, reaching about 545 to 790 tok/s total, with recall at 250K measured at 31/32. The author notes attention is approximate, so near-duplicate code lines can occasionally be mixed up.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
8 agents × 250K context, concurrently, on two 3090s (~600+ tok/s total)