{"slug": "introducing-interp-engine", "title": "Introducing: interp-engine 🚀🔎", "summary": "Decode Research has open-sourced interp-engine, a high-performance interpretability engine built from scratch to run production Neuronpedia workloads, including Jacobian Lens, NLAs, circuit tracing, activations, and steering. The v1 release supports 34 standardized hook points and claims to run Jacobian Lens for Deepseek V4 Flash about 7x faster than Transformers while producing the same outputs. The engine is vLLM-based with enhancements for reading and writing to multiple points during generation, and includes features like a visualizer, GPU Sizer, and an AI helpbot named 'Ask Riz'.", "body_md": "*This post is human-written and not AI generated or edited.*\n\n[interp-engine](https://interp-engine.org) is a high-performance interpretability engine built from scratch to run all our production Neuronpedia work, including Jacobian Lens, NLAs, circuit tracing, activations, and steering. We're open sourcing it for you (and your agents) to accelerate your interp research and workloads.\n\nThis is the v1 of interp-engine - we expect to improve performance significantly over the coming months.\n\nGet started at [interp-engine.org](https://interp-engine.org).\n\n[docs](https://www.interp-engine.org/docs) | [model points + viz](https://www.interp-engine.org/?arch=DeepseekV4ForCausalLM) | [github](https://github.com/decoderesearch/interp-engine) | [gpu-sizer](https://www.interp-engine.org/sizer/Qwen/Qwen3.6-27B)\n\n`interp-engine`\n\nlets you to do production-level interpretability work and research without compromising on speed and quality. In the race to better understand AIs as their capabilities increase, being able to observe and steer the latest LLMs at the same speed on comparable hardware should be table stakes.\n\nWe built interp-engine from scratch with these goals:\n\nBelow, we'll go over some of the highlights of interp-engine.\n\nIn the speed comparison below, we pit interp-engine against Transformers on running Jacobian Lens for Deepseek V4 Flash, with interp-engine completing the generation ~7x faster, while getting the same outputs.\nHow does it do this? `interp-engine`\n\nis vLLM by default, and carefully adds enhancements to support reading to and writing from 34 points while generating at near-production speeds (standard vLLM only supports residual stream), including new/novel points like multi-residual-stream.\n\nFull benchmark results across multiple models are at [results-latest.md](https://github.com/decoderesearch/interp-engine/blob/main/benchmarks/results-latest.md).\n\nThe increase in performance allows us to increase rate limits on Neuronepdia, host larger models, and increase max number of tokens in each application.\n\nStandardized points helps researchers compare apples-to-apples when doing investigations across models, and makes production interpretability deployments significantly simpler - no special casing for each model.\n\n`interp-engine`\n\nsupports 34 standardized hook points, and ships an intuitive visualizer of major model architectures, so you can see where the point you're grabbing is, how it's derived, and instantly copy a snippet for loading it - like a \"cheat sheet\" for model interp (from [interp-engine.org](https://interp-engine.org)):\n\nYou can also use \"Compare\" to compare two model architectures' differences. Here, we compare [Gemma 2 vs Gemma 3](https://www.interp-engine.org/?arch=Gemma2ForCausalLM&vs=Gemma3ForCausalLM):\n\n`interp-engine`\n\nis designed to be simple and 'ergonomic' for developers and ships with examples-first documentation at [interp-engine.org/docs](https://www.interp-engine.org/docs). It also contains extensive Markdown docs in its [repo](https://github.com/decoderesearch/interp-engine).\n\nTo get direct answers quicky, you can also just \"Ask Riz\" - an AI helpbot that knows all the interp-engine docs, on the navbar of interp-engine.org:\n\nTo avoid the headaches of finding the correct GPU configs to run in production that balances performance, GPU cost, and context sizes without OOMing, interp-engine also has a \"GPU Sizer\" feature - which intuitively guides you to the right GPU types to deploy on depending on your needs.\n\nAs an example: Here we pick Qwen 3.6-27B, then choose interp-engine's vllm-static backend for the fastest performance, at the cost of higher VRAM. We tell gpu-sizer that we need access to both resid_post and attn points. In the results, GPU-sizer automatically chooses the smallest single GPU config that will fit our desired characteristics, and shows us the visual breakdown of VRAM. It also tells us useful performance expectations like KV cache size, concurrent requests, etc. Finally, we copy the code that does this loading.\n\nGPU-sizer is also available [via API](https://www.interp-engine.org/docs/gpu-sizer-api), and works for your private models as well - just supply your HF token under \"Override HF Token\".\n\nFor both researchers and production deployments, nothing is more important than accuracy. To ensure that interp-engine's outputs are correct, we validated 50+ models by running (and will periodically re-run) comparisons against other interpretability engines, as well as HF Transformers. During the course of verifying our outputs, we discovered and helped resolve multiple incorrect outputs in other engines.\n\nFull comparision / validation suite, including scripts that you can use to re-run validations is available in the interp-engine repo under [interp-engine/validator](https://github.com/decoderesearch/interp-engine/tree/main/validator).\n\n`interp-engine`\n\nwas built with carefully directed and validated AI assistance. We expect that you'll use interp-engine via your AI/agents as well - it should be simple for you to simply append \"use the interp-engine.org library\" to your prompts and have it \"just work\". Let us know how it goes!\n\nWe expect to continue maintaining interp-engine to support new models as they come in, and also will continue to refine and improve its performance, as we rely on it for our core Neuronpedia and future platforms.\n\nAs always, please [contact us](https://www.neuronpedia.org/contact) with your questions, feedback, and suggestions.", "url": "https://wpnews.pro/news/introducing-interp-engine", "canonical_source": "https://www.neuronpedia.org/blog/interp-engine", "published_at": "2026-08-31 00:00:00+00:00", "updated_at": "2026-09-02 03:22:52.281445+00:00", "lang": "en", "topics": ["ai-research", "ai-tools", "ai-infrastructure", "machine-learning"], "entities": ["Decode Research", "interp-engine", "Neuronpedia", "Deepseek V4 Flash", "vLLM", "Transformers", "Qwen 3.6-27B", "Gemma 2"], "alternates": {"html": "https://wpnews.pro/news/introducing-interp-engine", "markdown": "https://wpnews.pro/news/introducing-interp-engine.md", "text": "https://wpnews.pro/news/introducing-interp-engine.txt", "jsonld": "https://wpnews.pro/news/introducing-interp-engine.jsonld"}}