Forward Deployed Engineer

The code reviewer that got 2x faster and measures bugs in compliments. Greptile shipped v5 claiming it catches more bugs, with higher precision, and 2x faster. The speed claim is exact: 5:04 to 2:25 median review time is 2.097x, measured by A/B test across more than a million live pull requests, which is a better design than most launch posts manage. The other two claims rest on three published numbers and none of them measures bug detection. Greptile maintains a public, reproducible bug-detection benchmark. The v5 post does not use it. Links & resources. | Resource | Link | | Announcement | @dakshgup | | The post | Greptile v5 | | Their bug-detection benchmark | greptile.com/benchmarks | | Previous version | Greptile v4 and new pricing | | Related research | Models are worse at reviewing their own code | Greptile announced v5 on August 5 with a three-part claim: "it catches more bugs, with higher precision, and 2x faster (~2 minutes per review)." One of those three is verifiable from the post, checks out to two decimal places, and is genuinely impressive. The other two rest on three published numbers, and none of the three measures bugs. The speed claim is exact. The post gives before and after directly: Read those as minutes and seconds, which the tweet's "~2 minutes per review" confirms is the intended unit, and 5:04 is 304 seconds against 145. That is 2.097x, so "2x faster" is if anything understated. The word "seconds" in that sentence is a slip, since 5:04 seconds is not a duration, but the numbers behind it are consistent with the tweet and with each other. Worth saying how they got it, because the method is better than the industry norm: Live production data across more than a million pull requests, run as an A/B against the previous version, is a stronger design than almost anything you will see in a launch post. Most releases in this category ship with a static benchmark of a few dozen cases and no control. What the other two numbers actually measure. Here are the remaining metrics, both from the post: Their arithmetic is right. 0.342 divided by 0.266 is 1.2857, so 28.6% is exact, and the 52 to 66 move is 14 points or 26.9% in relative terms. Now read them against the claim they are supporting. The tweet says v5 "catches more bugs, with higher precision." Precision, in review tooling, is the share of flagged issues that are real. Neither of these numbers is that. Comments addressed counts how often an author acted on a comment. That is a reasonable proxy for precision and I think it is the strongest number in the post, but it moves for other reasons too: a comment can be addressed because it was clearer, better placed, or arrived before the author moved on, all of which v5 also changed. Positive replies per PR counts how often a human typed something like "nice catch." That measures sentiment. It is also worth noticing where the level sits rather than only the delta: 0.342 per pull request means that after the improvement, roughly one PR in three draws a positive reply. Neither number can rise or fall in a way that distinguishes "found more real bugs" from "annoyed people less." A false-positive rate would. So would a bug-catch count. The chart makes the substitution explicit. I did not have to infer any of that, because the figure at the top of this post is Greptile's own and it labels its three panels like this: (A) BUGS CAUGHT · POSITIVE REPLIES / PR (B) PRECISION · COMMENTS ADDRESSED EXACTLY (C) SPEED · MEDIAN REVIEW TIME Each panel names a claim and then names the number standing in for it. Panel A is headed bugs caught and plots positive replies per pull request. Panel B is headed precision and plots the share of comments an author addressed. Only panel C measures the thing in its own heading. I find this more disarming than annoying. A chart that quietly plotted sentiment under a "bugs caught" axis would be the bad version. This one prints the proxy next to the claim in the same weight of type, which means anybody reading the figure can see the substitution being made and decide what to do about it. It is a strange kind of honesty, and it is honesty. They have a bug-detection benchmark. It is not in this post. This is the part I did not expect. Greptile publishes a benchmark page that measures exactly the thing the tweet claims: It reports bug detection by severity, and the page says the pull requests "come from public, verifiable repositories, so you can inspect the sources and reproduce the runs on your own." That page is dated 2025 and describes five tools across 50 bugs. It predates v5, and the v5 announcement does not reference it or add a v5 column to it. So the company that maintains a public, reproducible bug-detection benchmark announced a bug-detection improvement without running it. I want to be fair about the tradeoff, because it is real and it cuts both ways. A million live pull requests is a far larger and more representative sample than 50 curated bugs, and A/B data from production tells you things a static benchmark never will. But the static benchmark is the one that can measure precision, and it is the one a reader can reproduce. The v5 evidence is bigger and less specific. The benchmark is smaller and answers the actual question. What changed architecturally. The mechanism is stated plainly and it explains the latency result: Because the agents run in parallel, wall-clock time falls, and because each one holds a single hypothesis, each can search deeper before giving up. That is a coherent story for both halves of the claim, and the halving of median review time is consistent with it. It also suggests why precision might genuinely improve without anyone measuring it: an agent scoped to one hypothesis has less room to pattern-match its way into a comment it cannot support. That is a plausible mechanism, not evidence, and I am flagging it as the former. What to do with this. If review latency is your bottleneck, this is a real and verified improvement. 304 seconds to 145 on median, measured in production against a control. If false positives are your bottleneck, the post does not answer you. The closest available signal is that 66% of comments now get addressed against 52% before, which is a proxy and a good one, and it is not a precision figure. And if you want the precision number, the benchmark that would produce it is already built, public and reproducible on their own site. Adding a v5 column to it is a smaller job than the rewrite they just shipped.

New NVIDIA Nemotron 3 Super delivers 5x higher throughput for agentic AI. A new, open, 120-billion-parameter hybrid mixture-of-experts model optimized for NVIDIA Blackwell addresses the costs of long thinking and context explosion that slow autonomous agent workflows. Launched today, NVIDIA Nemotron 3 Super is a 120-billion-parameter open model with 12 billion active parameters designed to run complex agentic AI systems at scale. Available now, the model combines advanced reasoning capabilities to efficiently complete tasks with high accuracy for autonomous agents. AI-Native Companies: Perplexity offers its users access to Nemotron 3 Super for search and as one of 20 orchestrated models in Computer. Companies offering software development agents like CodeRabbit, Factory and Greptile are integrating the model into their AI agents along with proprietary models to achieve higher accuracy at lower cost. And life sciences and frontier AI organizations like Edison Scientific and Lila Sciences will power their agents for deep literature search, data science and molecular understanding. Enterprise Software Platforms: Industry leaders such as Amdocs, Palantir, Cadence, Dassault Systèmes and Siemens are deploying and customizing the model to automate workflows in telecom, cybersecurity, semiconductor design and manufacturing. As companies move beyond chatbots and into multi-agent applications, they encounter two constraints. The first is context explosion. Multi-agent workflows generate up to 15x more tokens than standard chat because each interaction requires resending full histories, including tool outputs and intermediate reasoning. Over long tasks, this volume of context increases costs and can lead to goal drift, where agents lose alignment with the original objective. The second is the thinking tax. Complex agents must reason at every step, but using large models for every subtask makes multi-agent applications too expensive and sluggish for practical applications. Nemotron 3 Super has a 1-million-token context window, allowing agents to retain full workflow state in memory and preventing goal drift. Nemotron 3 Super has set new standards, claiming the top spot on Artificial Analysis for efficiency and openness with leading accuracy among models of the same size. The model also powers the NVIDIA AI-Q research agent to the No. 1 position on DeepResearch Bench and DeepResearch Bench II leaderboards, benchmarks that measure an AI system's ability to conduct thorough, multistep research across large document sets while maintaining reasoning coherence. Hybrid Architecture. Nemotron 3 Super uses a hybrid mixture-of-experts (MoE) architecture that combines three major innovations to deliver up to 5x higher throughput and up to 2x higher accuracy than the previous Nemotron Super model. * Hybrid Architecture: Mamba layers deliver 4x higher memory and compute efficiency, while transformer layers drive advanced reasoning. * MoE: Only 12 billion of its 120 billion parameters are active at inference. * Latent MoE: A new technique that improves accuracy by activating four expert specialists for the cost of one to generate the next token at inference. * Multi-Token Prediction: Predicts multiple future words simultaneously, resulting in 3x faster inference. On the NVIDIA Blackwell platform, the model runs in NVFP4 precision. That cuts memory requirements and pushes inference up to 4x faster than FP8 on NVIDIA Hopper, with no loss in accuracy. Open weights, data and recipes. NVIDIA is releasing Nemotron 3 Super with open weights under a permissive license. Developers can deploy and customize it on workstations, in data centers or in the cloud. The model was trained on synthetic data generated using frontier reasoning models. NVIDIA is publishing the complete methodology, including over 10 trillion tokens of pre- and post-training datasets, 15 training environments for reinforcement learning and evaluation recipes. Researchers can further use the NVIDIA NeMo platform to fine-tune the model or build their own. Use in agentic systems. Nemotron 3 Super is designed to handle complex subtasks inside a multi-agent system. A software development agent can load an entire codebase into context at once, enabling end-to-end code generation and debugging without document segmentation. In financial analysis it can load thousands of pages of reports into memory, eliminating the need to re-reason across long conversations, which improves efficiency. Nemotron 3 Super has high-accuracy tool calling that ensures autonomous agents reliably navigate massive function libraries to prevent execution errors in high-stakes environments, like autonomous security orchestration in cybersecurity. Availability. NVIDIA Nemotron 3 Super, part of the Nemotron 3 family, can be accessed at build.nvidia.com, Perplexity, OpenRouter and Hugging Face. Dell Technologies is bringing the model to the Dell Enterprise Hub on Hugging Face, optimized for on-premise deployment on the Dell AI Factory, advancing multi-agent AI workflows. HPE is also bringing NVIDIA Nemotron to its agents hub to help ensure scalable enterprise adoption of agentic AI. Enterprises and developers can deploy the model through several partners: * Cloud Service Providers: Google Cloud's Vertex AI and Oracle Cloud Infrastructure, and coming soon to Amazon Web Services through Amazon Bedrock as well as Microsoft Azure. * NVIDIA Cloud Partners: Coreweave, Crusoe, Nebius and Together AI. * Inference Service Providers: Baseten, CloudFlare, DeepInfra, Fireworks AI, Inference.net, Lightning AI, Modal and FriendliAI. * Data Platforms and Services: Distyl, Dataiku, DataRobot, Deloitte, EY and Tata Consultancy Services. The model is packaged as an NVIDIA NIM microservice, allowing deployment from on-premises systems to the cloud. Stay up to date on agentic AI, NVIDIA Nemotron and more by subscribing to NVIDIA AI news, joining the community, and following NVIDIA AI on LinkedIn, Instagram, X and Facebook.