StartLux Interview: A 27B Local Model Rivals DeepSeek Giants via AutoResearch and Local RSI
机器之心 · wechat · 2026-09-16
- Context: 25 Fields medalists published a letter against AI labs benchmark-gaming famous math problems, while OpenAI claimed an internal model with 10,000 agents partially proved the Navier-Stokes problem in 88 hours. In CAICT's report, StartLux's 27B local model scored 39.25 on the MCP benchmark, ranking second—beating 284B DeepSeek-V4-Flash and within 1 point of far larger models.
- Key numbers: built on Qwen3.6-27B with essentially unchanged architecture, it gains +5.34 points entirely from post-training; 70% of the research pipeline involves AI (over 95% if only mechanical execution counts).
- AutoResearch method: AI models diagnose failures, generate improvement hypotheses, auto-build training data, launch training, rerun evals, and check for regressions; humans keep research goals, eval criteria, and key tradeoffs. The team stresses heterogeneous multi-model collaboration rather than distillation—"ground truth" comes from real environments and independent evals.
- Quantization findings: across 1,209 tasks on four benchmarks, Q4/Q6/Q8 stay within -0.1+0.2 points of BF16 for Qwen (Q2 drops 1.3); the 16GB-VRAM feasibility comes from joint quantization-plus-inference optimization.
- Capability reallocation: one agent-focused post-training round lifted GPQA from 83.33 to 90.40 while GAIA fell from 57.57 to 45.70. The CTO argues exploration freedom and self-modification permission must be separated, and that local models—not a single strongest model—are the path to AGI.
More from Infra
- OpenAI details scaling to 70M requests/sec over 500PB across nearly 40 regions — AxSaucedo · 2026-09-16
- NVIDIA's FlashDreams Gives Autoregressive Video and World Models an Inference Runtime — techNmak · 2026-09-16
- SK Hynix and Intel in Talks to Make Memory Chips in the US — zephyr_z9 · 2026-09-16
- Japan's chip exports catch up to autos as AI boom lifts the cycle — PAstynome · 2026-09-16
- Dev rebuilds Radish: a Redis server running entirely inside a Cloudflare Durable Object — whoiskatrin · 2026-09-16
- Anthropic lands its first Australian datacentre, a $31.9bn project in western Queensland — nordicinst · 2026-09-16