AI Agents and Developer Infrastructure #105
Today's Letter
Meta, Muse Glimmer 30B Local Agent Model Released
- Meta released Muse Glimmer, a 30B-parameter open-weight dense model for long-running local agent workflows
- The model provides a 120K+ context window for multi-step tool use and sustained sessions
- Its dense architecture activates all parameters per token, targeting predictable latency and long-context coherence
- NVIDIA supports local inference on GeForce RTX 5090, DGX Spark, DGX Station, and Jetson
- Deployment options include NVIDIA NIM containers, SGLang, vLLM, and the NemoClaw agent harness
- Developers can fine-tune the model with SFT and LoRA through NVIDIA NeMo AutoModel
Source: developer.nvidia.com
Atlassian StreamHub shifts from Kinesis to Kafka
- According to the report, Atlassian moved StreamHub from Amazon Kinesis to AWS MSK-based Kafka as event volume grew
- StreamHub reportedly ingests more than 150 billion events and delivers more than 225 billion events per day
- Average throughput is about 1.68 million events per second, with peaks above 3.2 million
- Kafka Tiered Storage moves older log segments to Amazon S3, reducing long-term EBS requirements and isolating backfills from real-time consumers
- The operating model adds broker-level capacity planning, deliberate network and disk headroom, quotas, rate limits, and quarantine paths
- Multiple Kafka shards and a separate failover cluster provide recovery options when the managed service control plane is impaired
- Reported constraints include broker network and EBS limits, S3 request bursts, offload lag, storage-expansion cooldowns, and slow rebalancing
Source: news.hada.io
More: atlassian.com
Jocoletter curates AI, software, and product trends for developers and builders.
#Atlassian #Meta