AI developer tools · What shipped

NeMo Gym v0.2.0 released alongside NVIDIA Nemotron 3 Super model, open sourcing 17 new RL training environments spanning coding (Text to SQL, SWE RL), math (Lean4 proofs), science (Aviary, NewtonBench), reasoning (ARC-AGI), agentic tasks, and safety benchmarks. New features include local vLLM model serving with end-to-end rollout collection, PyPI installation support, 5 new agent servers, environment library integrations (Aviary, Reasoning Gym, Verifiers), dry run mode for config validation, per-task pass rate profiling, and end-to-end training recipes for Nemotron 3 Nano on single-GPU and multi-node setups.

MemPalace, a local-first AI memory tool, achieved 96.6% retrieval recall (R@5) on the LongMemEval benchmark using raw semantic search with zero API calls and no LLM requirement. The system stores verbatim conversation history with structured indexing (people and projects as "wings", topics as "rooms") and supports pluggable backends including ChromaDB, Qdrant, Milvus, and pgvector. Additional benchmarks show 98.4% recall with hybrid retrieval (keyword and temporal boosting) and ≥99% with LLM reranking, while other datasets (LoCoMo, ConvoMem, MemBench) demonstrate 60.3%–92.9% recall depending on task complexity. The tool integrates with Claude Code, Cursor IDE, and other MCP-compatible clients through 45 MCP tools for palace operations, knowledge graphs, and agent coordination, with all computation remaining local by default and no cloud dependency for core functionality.

Parallel released benchmark results comparing its web search and research APIs against competitors (Exa, Tavily) across three datasets: SimpleQA Verified (1,000-question benchmark from Google DeepMind), BrowseComp (1,266-question OpenAI benchmark for complex web browsing), and WideSearch (200-task ByteDance benchmark for structured data collection). Parallel's Advanced tier achieved 97% accuracy on SimpleQA Verified at $28.30 CPM, while its Fast tier reached 94% at $2 CPM; on BrowseComp, Parallel Advanced scored 74% accuracy at $399 CPM compared to competitors' lower scores at similar or higher costs. Testing occurred September 9, 2026. Parallel also announced a suite of developer APIs including Search, Extract, Responses, Task, FindAll, and Monitor, all powered by its proprietary Web Index updated with millions of pages daily, designed for AI agents with per-request pricing starting at $1 per 1,000 requests.

Flatkey, an AI infrastructure platform, raised $10 million in Series A funding and surpassed 10,000 developers within two months of launch. The platform provides developers access to over 100 official AI models and 1,000 tools through a single API key, supporting models from OpenAI, Anthropic, Google, and DeepSeek. Flatkey offers pricing at approximately 80% of official list prices and functions as a drop-in replacement for OpenAI-compatible clients. The company plans to expand its model and tool offerings and scale its infrastructure to route developer traffic.