2 hrs ago
Nvidia-Backed Reflection Unveils Open-Weight AI Model Beam
Reflection AI has introduced an AI model called Beam.
The company wants it to compete with less expensive models from China.
Beam is designed to help with coding and tasks where AI systems carry out actions.
Reflection says Beam performs competitively with Z.ai’s GLM-5.2 and is getting closer to Qwen3.8-Max in those areas.
Beam has 501 billion parameters in total.
Parameters are parts of a model that help it process information.
For each task, Beam uses only 23 billion of them.
Reflection says using only part of the model can make it faster and cheaper to run.
Reflection AI introduced its first open-weight model, Beam, on Monday.
The startup said Beam is designed to compete with lower-cost Chinese models in coding and agentic tasks.
Reflection said Beam is competitive with Z.ai’s GLM-5.2 and is closing in on Qwen3.8-Max for those tasks.
Beam has 501 billion total parameters, according to Reflection.
The model activates 23 billion parameters per task, which Reflection says can make it faster and cheaper to run.
- Who
- Reflection AI, a startup backed by Nvidia.
- What
- It debuted Beam, its first open-weight AI model, to compete in coding and agentic tasks.
- Where
- When
- Monday; the article does not specify a date.
- Why
- Reflection aims to compete with lower-cost Chinese AI models.
Reflection’s claims
Independent verification
Beam’s comparative performance
Reflection’s claims
Reflection says Beam is competitive with Z.ai’s GLM-5.2 and is closing in on Qwen3.8-Max for coding and agentic tasks.
Independent verification
The article provides no independent assessment or response from the makers of the comparison models.
Efficiency and cost
Reflection’s claims
Reflection says activating only 23 billion of Beam’s 501 billion parameters per task can make it faster and cheaper to run.
Independent verification
The article provides no independent measurements of Beam’s speed or operating cost.
Key facts
- Model
- Beam, Reflection AI’s first open-weight model
- Launch
- Debuted on Monday; no date specified
- Total parameters
- 501 billion
- Parameters activated per task
- 23 billion
- Target tasks
- Coding and agentic tasks
- Comparison models
- Z.ai’s GLM-5.2 and Qwen3.8-Max
Sources
Related news
Alibaba unveils AI chip, plans trillion-parameter Qwen models
DeepSeek Narrows AI Performance Gap With US Rivals to 3%
OpenAI and Anthropic Release Cheaper Models Amid AI Slowdown Calls
GPT-6.1 Sol Offers Lower-Cost Alternative To Flagship GPT-6 Astra
OpenAI Expands GPT-6 Lineup With Cheaper Sol and Luna Models
