Amazon open-sources Strands Decider 2B decision model for local deployment
10/03/2026 — 10/03, 18:56·1 sources·1 reports
Story overview
On October 1 local time, the Strands Agents team at Amazon announced Strands Decider 2B, a decision-making model. The release was reported by IT之家 on October 3.
The model is open-sourced on GitHub, with weights available on Hugging Face, and it can run locally on CPU or GPU. According to the report, Strands Decider 2B is built on a pretrained Qwen3.5-2B "trunk." The text-generation head of the original language model is replaced by a pointer "head" that performs a scoring function. That new head is small, with just over one million parameters in total, while the trunk is fine-tuned through a rank-16 LoRA adapter.
On the public JevBench dataset, Strands Decider 2B is reported to show solid accuracy and calibration. It ranks third among 2B-class models, ahead of every competitor strictly at or below 2B. When run locally on commonly available hardware, the model posts a median latency of 153ms on an RTX 3090 for small decision-making tasks. The source excerpt also cuts off a second performance figure, so that number cannot be confirmed from the material available. No further developments beyond the initial announcement and the October 3 report have been disclosed.
AI-generated from 1 reports · updated 2 hours ago
Latest turnAmazon's Strands Agents team has released Strands Decider 2B, a decision model with weights on Hugging Face that runs locally on CPU or GPU. It pairs a Qwen3.5-2B backbone, fine-tuned with a rank-16 LoRA adapter, with a small scoring head of just over one million parameters that replaces the original text-generation head. It ranks 3rd among 2B-class models on JevBench and hits a 153ms median latency on small decision tasks on an RTX 3090.
- Heat index
- 91
- Sources
- 1
- Reports
- 1
- First seen
- 4 hours ago
Reports on this story headlines open the original
Amazon's Strands Agents team has released Strands Decider 2B, a decision model with weights on Hugging Face that runs locally on CPU or GPU. It pairs a Qwen3.5-2B backbone, fine-tuned with a rank-16 LoRA adapter, with a small scoring head of just over one million parameters that replaces the original text-generation head. It ranks 3rd among 2B-class models on JevBench and hits a 153ms median latency on small decision tasks on an RTX 3090.
IT之家AI score 85
Other stories people are talking about
- 529RisingApple tightens macOS Full Disk Access over AI agent risks8 sources
- 453NVIDIA launches 64GB DGX Spark desktop AI computer at $4,9998 sources
- 334SurgeClaude Opus 5.5 and GPT-6 Sol: comparing cost per task4 sources
- 327Meta open-sources Muse Gadgets firmware and SDK5 sources
- 219Microsoft AI releases MAI-Transcribe-2-Streaming real-time transcription model4 sources
- 192Amazon Weighs Moving $8 Billion of Nvidia Chips into a Financing Vehicle3 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion