cactus-compute/needle

This page introduces an external open-source repository. It is not an HDATF product.

What it does

Needle 2 is an open 45M-parameter model for tool calling, device use and structured extraction, packed into a single 14MB engine. The Python package covers inference, LoRA fine-tuning and export, returning JSON tool calls with a confidence score.

How it helps ATF

Its pattern of acting above a confidence threshold and escalating below it could be a reference for Harness when weighing whether a result can be accepted or needs escalation. Otherwise it is only loosely related to our products.

License

Apache-2.0 Permissive, with a patent grant. Commercial use is allowed; keep the notices and state your changes.

More in this category

• FlashML-org/FreeToken
  A serving engine for running large open-weight Mixture-of-Experts models on personal hardware by splitting work across GPUs, CPUs and memory. It exposes Anthropic- and OpenAI-compatible APIs so coding agents can connect.
• unslothai/unsloth
  Unsloth is a desktop app for running and training LLM, diffusion, embedding and audio models, including GGUF and MLX formats. Local models can also be used with Claude Code, Codex and MCP, together with web search and RAG.
• sgl-project/sglang
  A serving framework for large language models and multimodal models. It serves open models such as DeepSeek, Qwen, Llama and GLM, and also covers diffusion models for image and video generation.
• FareedKhan-dev/kimi-k3-in-c
  A portable C99 engine that runs Kimi K3 inference on one CPU with no GPU, BLAS or framework, streaming the model from disk. The readme reports an 8.24 GB peak memory use, with more RAM only adding speed.
• jingyaogong/minimind
  An educational project that implements a small language model of about 64M parameters from scratch in native PyTorch. Its code covers the full training chain, such as pretraining, SFT, LoRA, RLHF, tool use and distillation.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.