IBM released Granite Speech 5.0 2B, an Apache 2.0 speech recognition and translation model ranking second on the Open ASR Leaderboard with 5.68% average English word error rate, alongside updated Turbo CTC and 4.1 2B Plus variants (MarkTechPost). The model is designed for multilingual transcription and translation tasks.
Recursive released Flint, an Apache 2.0 open-weight 40B MoE coding model with 2.5B active parameters optimized for local deployment and agent tasks (MarkTechPost). The model reports 74.2% on SWE-bench Verified and achieves 125+ tokens per second on an M4 Pro MacBook Pro, supporting on-device code generation workflows. Separately, Vercel open-sourced durable-harness, a package that logs model and tool calls to enable coding agent sessions to resume after crashes or timeouts across deployment platforms (MarkTechPost). Additionally, Perplexity and NVIDIA researchers introduced JOLT, a post-training method that matched GRPO accuracy on MATH while consuming 6.5x fewer processed training tokens (MarkTechPost).