AI developer tools · What shipped

DeepSeek released DeepSeek-V4.1-Flash-1017, delivering improved benchmark performance and faster task completion compared to the previous version with reduced output length and cost. The model is optimized for DeepSeek Harness v0.2.1-rc.1 and other developer tools, and uses reinforcement learning with self-feedback to enable end-to-end project development from game creation to 3D modeling. Additionally, DeepSeek's Search & Fetch APIs are now publicly available at $0.75 per 1,000 calls, with free usage through server-side tools in DeepSeek Harness, Claude Code, and Codex until billing begins October 31, 2026. The model is available open-source on Hugging Face and ModelScope.

Cactus Compute released Needle 3, a foundation model (8-29 MB) optimized for mobile, wearable, and edge devices, featuring a Laddered Simple Attention Network architecture. Needle 3 excels at tool calling and structured extraction with exact-match accuracy benchmarks on full test splits, matches or beats models 10x its size on mobile tool use cases, and includes LoRA fine-tuning capabilities for custom deployments. The company also released Whistle, a 16.9 MB speech-to-text model that is 8.6x smaller than Whisper base with faster latency, supporting seven languages with benchmarked performance on LibriSpeech, SPGISpeech, and FLEURS datasets. Both models ship as single binary containers with unified C++ runtime, available via pip/npm with weights on Hugging Face and interactive demos at cactuscompute.com.

DeepSeek released V4.1-Flash-1017 on October 17, 2026, a same-price checkpoint update to its deepseek-flash API model with MIT-licensed weights. The model gains 11.1 points on Terminal-Bench 4.0 and 8.8 on ProgramBench over the September version, with the largest improvements on long-horizon agent work. According to DeepSeek's benchmarks, it beats Opus 5 on DeepSWE v1.1 (77.0 vs 74.0), Automation-Bench, Agents' Last Exam, and Apex Bench, though Opus 5 still leads on HLE, Terminal-Bench 4.0, ProgramBench, and NL2Repo-Bench. The improvements come from reinforcement learning with self-feedback aimed at end-to-end project development, with the model architecture remaining a 552B-parameter MoE with a causal encoder-decoder layout. Pricing remains unchanged at $0.30 per million input tokens (cache miss) and $1.20 per million output tokens at peak hours, with off-peak rates at half price. The weights are available on Hugging Face and ModelScope with a 1M token context window and 384K maximum output.