whale (meat) lets you review what matters on a budget

Mon 20 July 2026 · 610 words · 4 min read

starting around March 2026, the bottleneck in many software engineering teams have move to reviewing AI generated code instead of writing every line by hand.

ai coding harnesses have gotten much better since then and many of the common mistakes like missing imports, null pointers that reviewers need to look carefully at have mostly been solved. the things that matter have moved to the systems design, business logic, and higher level considerations. LLMs are actually really great at catching the small details like proper error messages and comments way better than most developers including myself.

so i came across a neat tool called meat whiling reading this article which i find most agreeable, that uses a LLM to filter out the changes that do not matter and only show the critical lines that do matter a lot like architectural changes and business logic.

meat is written in Go, it works by using LLMs to decide what changes are important to review and filters out the rest. the compiler is responsible for the actual filtering, so the LLM does not write or summarize anything. it just proposes, and the compiler does the actual diffing.

the tool only supports OpenAI and Anthropic latest models by default. i had it scan one of my recent commits, around ~2200LOC. it filtered down the changes down to around ~650 lines from the 2200+ and it took a couple minutes and cost around $2 and over 1M tokens using GPT-5.6-Sol. that saves me around 2/3rds of the time by showing me only what matters.

however, for a small business, it seems like not something i would be able to afford to run dozens of times a day or as part of of some build pipeline. let us fix that and make it affordable for devs in the rest of the world.

3 days ago, Deepseek-V4-Flash-0731 was updated for general availability. it jumped from 7 points to 54.4 on DeepSWE benchmark and 82.7 on Terminal-Bench, most of these gains were from post training since deepseek-flash uses the same 245B parameter base model.

i figure this would be a good exercise to test Deepseek-V4-Flash new capabilities itself by seeing if it can build an integration for meat to use itself as a provider model, and then seeing how well Deepseek-V4-Flash does as an model for deciding what code changes were worth reviewing.

i had deepseek investigate and explore the initial project, identify the critical path and interesting parts like i do with every project i git clone, then i had it come up with a plan to add Deepseek as another model provider. one thing i had to confirm is that v4 flash was able to handle tool calling as the project depends on, so i had Pi write a probing script to query Deepseek-V4-Flash with tool use functions and see if it choose the correct tools. once this was confirmed to work, the rest of the integration was straightforward. Deepseek API supports the OpenAI Chat Completions response format while OpenAI default models only supports the new Responses API, so we need to create a new Model provider for Deepseek.

i had Pi (using Deepseek-V4-Flash) write an implementation plan, which i then feed to Pi (still dsv4-flash) in another window for the full implementation in one-shot.

here is the results which i renamed whale https://github.com/zshanhui/whale

now with Deepseek-V4-Flash support, the review only cost a few cents instead of $2, and i can see myself running this a dozen times per day or as part of an automated workflow. i still need to do some more evals to make sure the judgement is still on par with GPT class models, but so far vibe checks out.

all credit goes to the original project https://github.com/boldsoftware/meat and creators.