ARC 1.0: A routing model for selecting models for agents

ARC, an adaptive routing controller uploaded by Rayline on 11 10, is based on Qwen3.5-0.8B. It reads agent conversations and outputs signals such as success probabilities for each model, token usage (uncached input, cache read/write, output), reasoning intensity benefits, personal sensitive information detection, and request classification. The application then combines these with its own model pool, pricing, and rules to determine which model should handle the next request. Two preset strategies are provided: Sage (V=4.5, prioritizing success) and Swift (V=0.1, prioritizing cost savings). Utility = V × predicted success rate − expected cost; a “local-first” threshold can also be set. Apache-2.0.

Original title: rayline-ai/ARC-1.0

ARC, an adaptive routing controller uploaded by Rayline on 11 10, is based on Qwen3.5-0.8B. It reads agent conversations and outputs signals such as success probabilities for each model, token usage (uncached input, cache read/write, output), reasoning intensity benefits, personal sensitive information detection, and request classification. The application then combines these with its own model pool, pricing, and rules to determine which model should handle the next request. Two preset strategies are provided: Sage (V=4.5, prioritizing success) and Swift (V=0.1, prioritizing cost savings). Utility = V × predicted success rate − expected cost; a “local-first” threshold can also be set. Apache-2.0.

Why it matters Prediction and pricing strategies are separated, so changing prices does not require retraining; officially claimed that when mixing local and cloud models, quality approaches top-tier closed-source models while costs drop significantly

Rising score 34 · Likes 42 · Downloads 1786

Reasons

  • The technical report states: When mixing local DeepSeek V4.1 Flash (on multiple DGX Sparks) with cloud models and setting the local-first threshold to 0.85, it retains 97.5% of Opus 5.5 quality across 6 benchmarks, reducing costs by 92.9%
  • ARC Sage compared to using only Opus 5.5: DeepSWE 68.7% vs. 69.9%, cost is 7%; however, Tau3 Banking scores 33.2% vs. 42.8%, showing a significant drop
  • On the internal Router Question Bench with 3,056 questions, the agreement rate with target answers is 56.2%, while Jev 1.14 is 32.7%

Usage

  • Hosted service: Integrate with existing proxies via docs.rayline.ai; model IDs are arc-sage-1.0 and arc-swift-1.0
  • Local decision-making: Configure the open-source Rayline CLI with routing.decisions: "local", reading the gguf/ and runtime/ directories in the repository; routing decisions are not sent to the Rayline cloud (local mode is a preview)
  • Installing the open-source vllm-arc plugin on vLLM 0.33.0 allows serving the encoder, but does not include routing strategies or decision APIs; prices are in runtime/prices.json and can be replaced with your own

Risks

  • Evaluations come from the company’s own technical reports and internal question banks; actual performance depends on your model pool, pricing, and requests
  • Local decision-making is a preview version, and the vllm-arc plugin is in 0.1.0 pre-release, so interfaces may change; full integration currently requires using the hosted API
  • Custom architecture requiring accompanying code. This entry has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates