StartLux Beats Jev in 3 Days: Inside Their Auto Research RSI Pipeline
Chinese AI startup StartLux open-sourced StartLux-Decision, a decision model that surpassed Jev on 31 of 38 benchmarks in just three days of development, powered by their Auto Research system that automates 95% of training workflows and 70% of research direction decisions, while pioneering a five-level RSI framework and a local-first intelligence architecture.
StartLux-Decision Dethrones Jev in Holiday Release
On September 30, during China's National Day holiday, StartLux (原点星辉) — a company founded just over four months earlier — open-sourced its decision model StartLux-Decision . The team went from direction-setting to first-round R&D validation in only three days . In the independent Decision Index 0.2.1 benchmark suite of 38 tasks, the 27B version outperformed the then-leading U.S. model Jev 1.13 on 31 tasks , achieving a composite score of 63.88 vs. 57.91 . By October 8, Hugging Face downloads exceeded 53,000 , and independent third-party testing on a 248-question suite confirmed that among locally deployable open-source models, only StartLux-Decision-27B approached the cloud-hosted Jev (total score 0.883 vs. 0.907), even surpassing Jev's raw accuracy when confidence gating was enabled.
Benchmark Breakdown & Model Capabilities
Strengths: Language understanding, retrieval & classification, tool use & automation — StartLux leads in the majority of these decision-critical categories.
Weakness: Knowledge & reasoning trails Jev (44.3 vs. 51.4).
Multimodal edge: Jev accepts only text + JSON; StartLux-Decision adds image input , enabling decisions directly from visual state.
Latency: 4B version answers three questions in 26 ms on a single H200 GPU (~10× faster than a human blink).
Footprint: 4B quantized weight is 4.48 GB (≈ a HD movie); quantized outputs match full-precision exactly on 231 public questions, proving it fits consumer hardware.
Real-World Demonstrations
HELURIO Kids Fit (GitHub project): 4B model runs on an M5 Pro MacBook, interprets a child's spoken request ("change to yellow T-shirt") in ~100 ms, triggers cloud try-on service. Developer chose 4B for optimal balance between spoken-language understanding and response speed.
StarCraft II: Model selects build orders and tactics, beats built-in Hard AI.
Chess: 35 wins out of 36 games against Jev.
Shopping task: Given "cheapest 8-pack AA batteries with free shipping," simultaneously verifies model, count, price, shipping — satisfies all constraints.
Auto Research: The Engine Behind 3-Day Iteration
StartLux's Auto Research system turns model development into a sustainable engineering pipeline. Key metrics from CTO Guo Quanwei:
Base-model swap + full post-training: ≤ 1 week , with >90% data & experience reuse.
Automation of "grunt work" (config generation, training launch, evaluation): >95%.
AI participation in "what to research next" (hypothesis generation, failure analysis, next-step selection): ~70%.
Concrete failure-driven loop example (finance task): Model gave wrong calculation without re-checking source data. System detected failure → multiple AI agents analyzed root cause & proposed fixes → auto-constructed corrective data → retrained → re-evaluated for regressions → promoted to next cycle.
Five-Level RSI Framework (Self-Assessed at Level 3)
Level 0 Self-correction: retry on error. Level 1 Memory/Experience: accumulate & reuse experience. Level 2 Auto-training: autonomous data generation, training launch, evaluation. Level 3 Auto-research: analyze failures, propose hypotheses, decide next steps. Level 4 True RSI: the ability to improve the improvement process itself — recursive feedback.
StartLux places itself at Level 3 , explicitly acknowledging it has not yet reached true recursive self-improvement. This disciplined taxonomy mirrors OpenAI's five-stage AGI roadmap in structural influence.
Local-First Intelligence Architecture
StartLux treats "local intelligence" as a full-stack system, not merely a compressed model. The architecture layers (pre-planned, not retrofitted) include:
Decision model (high-frequency judgments)
Agent & memory system (above)
Inference engine, quantization, hardware adaptation (below)
Industry validation arrived in September: Jev's rise, OpenAI's Decisions API, Cloudflare's Clef all confirmed the decision-model layer — a slot StartLux had already reserved in its blueprint. The next planned layer is Personal Agent .
Why Local Enables Faster RSI
Lower trial cost: Compact local models allow more experiments per GPU-hour. Karpathy's autoresearch achieves ~5 min/experiment; StartLux's scale enables similarly dense iteration.
Data locality: Each user's device generates a private evolution signal. Founder Chen Danian envisions 8 billion personalized models, each on its own trajectory, with updates gated by isolated sandbox + regression testing + instant rollback.
Strategic Moats & Outlook
Repeatable production line: Two models, six size variants, one-week base-model swap, 90%+ asset reuse in one month.
Pre-validated architecture: Blueprint anticipated industry-recognized capability layers before they emerged.
Honest RSI rubric: Public five-level standard with self-assessment at Level 3 — a rare discipline during peak hype.
StartLux's first consumer-facing local intelligence system is reportedly in development; its launch will test whether the engineering pipeline translates into end-user product experience.
}
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
