⚡

Developer 3s Key Decision Metrics

TL;DR Verdict
🏆 Benchmark Rank72s 必看
Arch: Qwen3.5-2B + LoRA r16 + 1M poi
Full Leaderboard→
💰 Cost & Plans比价
Arch: Qwen3.5-2B + LoRA r16 +
Compare Plans↗
🛠️ Ecosystem & IDERules
strands-agents
Setup IDE Rules↗

On 2026-10-01 AWS Strands Labs released Strands Decider 2B (strands-labs/strands-decider, Apache-2.0; weights HF hobson-v19): LM head replaced by an option-scoring head. JevBench public 167/231 ≈72.3%, Brier ~0.35; ~115ms median on RTX 3090. pip install strands-decider for routing, tool choice, and before_tool_call gates.

Key Takeaways

  • ✓Blog 2026-10-01; github.com/strands-labs/strands-decider Apache-2.0 (~140★); HF StrandsAgents/strands-decider-2B-hobson-v19
  • ✓Arch: Qwen3.5-2B + LoRA r16 + ~1M pointer head; noul/choice/score + confidence (no free text)
  • ✓JevBench public 167/231 (~72.3%), Brier ~0.348; 3rd of 33 in 2B class per authors
  • ✓Latency ~115ms median RTX 3090; ~153ms small tasks on M3 MacBook
  • ✓Try: pip install strands-decider; ask/serve; wire into Strands before_tool_call interventions
🧭

Finished reading? Explore benchmark rankings & pricing

Real-world SWE-bench scores & $20/mo vs API cost break-even calculator

🏆 Leaderboard→💰 Compare Plans↗
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
📝
Introducing Strands Decider 2B — Strands Agents Blog
strandsagents.com/blog/introducing-strands-decider/
↗
🐙
strands-labs/strands-decider (GitHub)
github.com/strands-labs/strands-decider
↗
🌐
StrandsAgents/strands-decider-2B-hobson-v19 (Hugging Face)
huggingface.co/StrandsAgents/strands-decider-2B-hobson-v19
↗
📝
TechCrunch: Amazon releases its own Jev clone
techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/
↗
🚀
Qwen3.5-2B
huggingface.co/Qwen/Qwen3.5-2B-Base
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points

Many agent steps are finite-choice decisions—routing, tool gating, groundedness—where full LLMs are slow/expensive and lack calibrated confidence. After TypeSafe Jev spotlighted decision/system-one models, AWS Strands Labs released Strands Decider 2B on 2026-10-01 for fast local experimentation and harness gates.

Architecture Highlights & Internals

Starts from Qwen3.5-2B, removes the LM head, adds a ~1M-parameter pointer head scoring option hidden states vs the answer position; torso uses rank-16 LoRA. Primitives: noul/choice/score with confidence. Code: github.com/strands-labs/strands-decider (Apache-2.0); checkpoint HF hobson-v19. Fits Strands before_tool_call interventions (Proceed/Deny/Confirm/Guide).

Authoritative Benchmarks & Measured Scores

Author-reported only: JevBench public 167/231 (~72.3%), Brier ~0.348; ~115ms median on RTX 3090; ~153ms small tasks on M3. Limitations: weak on long multi-step docs; question paraphrases may not flip answers; recalibrate thresholds on your traffic.

Developer Hands-on Guide

pip install strands-decider; ask/serve locally; load StrandsDeciderModel in Python; wire hybrid agents that leave hard calls to LLMs.