AI model and GPU infrastructure updates #140
Today's Letter
DeepSeek-V4.1-Flash-1017, minor model update

- DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash targeting reasoning, coding, agentic tasks, and visual benchmarks
- DeepSWE v1.1 score reaches 77.0, up from 74.2 in the previous model
- Terminal-Bench 4.0 improves to 42.3 from 31.2, while ProgramBench rises to 29.1 from 20.3
- Tool-based results include 83.0 on Chartography and 92.5 on BabyVision
- The model card lists 763B total parameters with sharded safetensors files totaling about 510GB
- Transformers, vLLM, SGLang, Docker Model Runner, and compatible quantization workflows are documented
- The release does not include a Jinja chat template and provides a self-contained Python prompt-encoding reference
Source: huggingface.co
More: startupfortune.com · shattered.io
NVIDIA Run:ai, GPU orchestration for Kubernetes
- NVIDIA Run:ai is an open-source AI workload and GPU orchestration platform for shared infrastructure on Kubernetes
- Supports development, training, and inference workloads across on-premises, cloud, and hybrid environments
- Combines GPU scheduling, resource management, workload orchestration, governance, and observability
- The control plane centralizes workload submission, resource management, cluster monitoring, and analytics
- Cluster services extend Kubernetes scheduling with GPU-aware workload management
- Users and external systems can connect through the web UI, CLI, and REST API
- The monorepo includes control-plane and cluster services, installer, CLI, MCP server, documentation, clients, and local development tooling
Source: github.com
More: techrepublic.com · aithority.com · neoteo.com
Jocoletter curates AI, software, and product trends for developers and builders.
#DeepSeek #NVIDIA