Spark-X2.6-1.7B Lands: 1M Context On-Device Agent Model Nearly Matches Its 4B Sibling
vllm_project · x · 2026-10-28
SparkLLM released Spark-X2.6-1.7B, a compact on-device agent model with 1M native context and controllable thinking (none/low/high). Despite less than half the parameters of its 4B sibling, it scores close: SWE Verified 59.8 vs 62.8, τ²-bench 85.3 vs 86.9 (both high thinking). vLLM offers Day-0 support via trust-remote-code, including vLLM-Ascend for NPU deployment; weights are on Hugging Face.
Related event: iFlytek Open-Sources Spark-X2.6-1.7B, Closing In on Its 4B Model(4 posts)→
More from Infra
- A forgotten supercomputer design: bare dies, gold post interconnects, 3D stacking, immersion cooling — jwt0625 · 2026-10-28
- Karpathy recommends a YouTube series building a CPU from a single transistor — blaizedsouza · 2026-10-28
- Idle GPUs burn cash: Neolabs lose money below ~60% spot utilization — FinanceYF5 · 2026-10-28
- Napkin math for GPU fit: a 7B model needs ~14GB at FP16, ~3.5GB at INT4 — cneuralnetwork · 2026-10-28
- Training on 1,000 GB300s costs $125-150M over 3 years and still trails frontier by 1-2 orders of magnitude — FinanceYF5 · 2026-10-28
- The hardest part of Neolab startups isn't training models — it's the math — FinanceYF5 · 2026-10-28