Phind

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Phind was a developer answer engine: a search box that retrieved documentation and forum threads, then drafted an explanation and a code snippet with links under it. Michael Royzen and Justin Wei started it in 2022. Y Combinator lists the company in its Summer 2022 batch, and PitchBook records an earlier name, Hello, and a San Francisco office. The product was never a general chatbot and never a hosted training platform. It was a narrow interface for people who would otherwise have had a dozen tabs open on a bug. That interface is gone. Secondary accounts of a January 2026 email from Royzen, including posts that circulated on Hacker News, agree on the operational dates even where they differ by a few days on when the note went out. Search stopped on 16 January 2026. Paid users were told to expect a prorated refund. Stored chats and saved searches could be downloaded until 30 January 2026, after which Phind said the data would be deleted. Y Combinator's company page has since marked Phind inactive. PitchBook's profile lists an "out of business" status dated 16 April 2026, which is a later database marker, not the day the site went dark. Do not send a new question to phind.com and expect the old product. The line in the navigation above, that no comparable persistent first-party agent is identified, is still the right comparison. Phind could see a question and, in the editor extension, some of the file you had open. It did not keep a long-lived workspace, run your tests overnight, or open pull requests the way later coding agents do. Closing the service removed even the search box. What survives in public is a handful of open weights on Hugging Face and the record of a product category that ChatGPT, Claude, Gemini, and Perplexity then absorbed.


Phind answered coding questions by searching the web and drafting a snippet. It shut the search service down on 16 January 2026, and it never became a persistent coding agent.


What the product actually did

A Phind query was retrieval-augmented generation aimed at technical pages. The system searched the live web, preferred documentation, repositories, and forum threads over general articles, and put the retrieved passages into the context of a language model that then wrote the answer. The citation was the product. A model that has only its training cut can describe an API that changed last month. A model that has just been handed the current docs is still able to misread them, but the reader can open the link. That is why developers treated Phind as closer to a better Stack Overflow page than to a pair programmer that lived in the repo. The optional codebase connection and the Visual Studio Code extension were the other half. You could ask about the file in front of you without pasting it into a browser. That is pair programming only in the weak sense: a suggestion attached to a buffer, not an agent with memory of a multi-day task. There was no public per-token API. Billing, while the site lasted, was a seat on the website and the extension. Secondary price lists from 2025 and 2026 do not agree with each other (figures around $17, $20, $30, and $70 a month all appear), and the pricing page is gone, so this article does not treat any of those numbers as a rate you can pay. The previous version of this page printed a long Visual Basic macro, generated as a sample, that colored spreadsheet words to match a list on another sheet. It is a fair specimen of the output and a fair specimen of the risk. The script invents sheet names ("Settings" and "Facts") and a column, uses a case-insensitive search on text it has already uppercased, and was not shown running in Excel. Copying it would have been an act of faith. Phind's useful answers were the ones a developer then compiled, ran, and checked against the cited page. The polished ones that only looked complete were the failure mode, and they looked the most complete. Late in the product's life the model behind the box kept changing, which is what a small team does when it cannot train a frontier model and cannot ignore the ones other people release. Users quoting Phind's own late-2025 notes said the "Fast" tier was based on GLM-4.5-Air and the "Large" tier on GLM 4.6, after earlier years in which the default had been Phind's own fine-tunes or a frontier API model. A Hacker News thread at the shutdown treated that switching as a sign the product had become a wrapper. The more careful reading is that the wrapper was always the business: search, citations, and a developer-shaped page. The weights were a means. When OpenAI, Anthropic, and Google shipped web search inside the same apps developers already paid for, the separate page had less to sell.

The open code models

The artifact that can still be inspected is Phind-CodeLlama-34B-v2, published on Hugging Face in 2023. The model card says it is a native fine-tune, not a LoRA, of an earlier Phind code model in the Meta Code Llama family, trained on another 1.5 billion tokens of programming instruction pairs. Phind reported 73.8 percent pass@1 on HumanEval and, at the time, called that the best open-model score. Training details on the card are unusually concrete for a startup release: sequence length 4096, DeepSpeed ZeRO stage 3, FlashAttention 2, and about 15 hours on 32 A100 80 GB GPUs. The card also says the set was instruction pairs rather than raw completion, and that the model handled Python plus C, C++, TypeScript, Java, and others. Those are the company's measurements. HumanEval is a few hundred short Python tasks. A high pass@1 meant the model was good at that exam, not that it could maintain a service. The episode that popularized the work, Latent Space in November 2023, advertised a nearby figure, 74.7 percent, against a much lower base Code Llama score, plus a longer context and very high tokens per second. Use the model card when you need a number you can point at, and treat the podcast as the narrative of how a tiny team fine-tuned rather than as a second benchmark. Later product names, Phind-70B and Phind-405B, were marketed as larger Llama-family code models inside the paid app. One 2026 writeup says a 70B fine-tune of Llama 3.3, with a 128K context, was posted to Hugging Face on the data-deletion date. Check the Hub before you assume a particular repository is still the weights Phind served in the app. Open weights outlive the company only if someone keeps the files up. Fine-tuning on a proprietary set of coding instructions was Phind's attempt at a moat that did not require a frontier training run. It worked as a 2023 story, because a 34B model that beat larger general models on one coding exam was news. It did not stay a moat. HumanEval saturated, the base models improved monthly, and the expensive part of the product became inference on whatever model the user expected that week, plus the search index in front of it. A subscription does not cover that cost once the comparison set is a frontier lab's own search-enabled coder. Community comments at the shutdown said this in plainer language: owning a fine-tune did not scale, and reselling someone else's API did not either.

Why it closed, and what that does and does not prove

Phind did not publish a postmortem that this page can quote as the cause. The email, as relayed by users, called the decision difficult and then gave dates. Everything else is interpretation, and some of it is sour. The timing is still a fact worth sitting with. PitchBook shows an early-stage venture deal on 3 December 2025, and contemporary accounts, including a Hacker News thread, put a round of about $10 million roughly six weeks before the lights went out. PitchBook's total capital raised for the company is about $10.9 million. Other writeups call the December deal a $10.4 million Series A. The amounts are close enough to say the company had just taken a conventional seed-or-A check, and not close enough to print one figure as audited. Readers on Hacker News offered the explanations you would expect, and they are not mutually exclusive. Frontier products had added the two things Phind was known for, code and live sources. A developer who already paid for ChatGPT or Claude did not need a third tab that called similar models. Perplexity had taken the general "answer engine" position with a much larger consumer audience. Usage was said, in some secondary reviews, to have fallen sharply from a 2024 peak. Those reviews do not show their traffic source, so the steep percentages should not be copied into a slide. Separately, long-time users complained that the team changed direction often, shipped features people did not want, and talked more than it shipped. That is a community argument about a small company's execution. It is not a measurement. The short notice is the part that generalizes past Phind. People had used the saved threads as a working notebook: the answer plus the links, kept next to a ticket. Two weeks to export, then deletion, is a hard lesson in where your notes live. A vendor's chat history is not your archive unless you can take it with you on an ordinary Tuesday, not only after a shutdown mail. The same lesson applies to any coding assistant that stores prompts, repository snippets, or customer code. Phind's zero-notice-adjacent ending does not mean those tools are unserious. It means a seat on a startup's website is a convenience, not a system of record. It also does not mean developer search was a bad idea. You began as a consumer answer site and survived by selling the retrieval API to other agents. The frontier labs kept search and folded it into the coder. Phind stayed the website. Once the website was no longer a better interface than the apps around it, there was nothing else. The navigation's remark about agents is the strategic version of that sentence. A persistent agent that lives in the repository, holds a branch, and can be pointed at a backlog is a different product from a cited answer. Phind never shipped the first one. By the time that product category was obvious, the company had already decided to stop.

Recent Developments

There is no current Phind product to update. From December 2025 through April 2026 the record is the round, the January shutdown, the deletion date, and the database status that followed. Royzen's Y Combinator profile, as described in 2026 writeups, lists him as previously the co-founder and chief executive of Phind and as the founder of a newer company, Standard Signal. That is his next listing, not a successor that runs phind.com. The Hub repositories from 2023 remain the technical primary source. Pricing pages, the VS Code workflow, and any "Phind 3" visualization launch from the December 2025 period are historical, and several review sites were still describing the product as live months after it was not. For a working substitute, match the job rather than the logo. Cited technical search is now a mode inside Perplexity, ChatGPT, Claude, and Gemini. In-editor completion and multi-file edits live in the pair-programming tools already linked from AI Software Engineering, including the frontier coding agents. None of them is Phind, and none of them inherits your deleted Phind threads. If a page or a video still shows a Pro plan and a daily search cap, it was written against a product that stopped taking queries on 16 January 2026.

References

  • Phind-CodeLlama-34B-v2 | Hugging Face The model card: 73.8 percent HumanEval pass@1, the extra 1.5 billion tokens, and the training setup. This is the best primary technical source the company left behind.
  • Phind.com is shutting down | Hacker News Developers reacting to the closure, the recent fundraise, and the sense that frontier search-and-code products had absorbed the niche. Commenters are not the company.
  • Phind company profile | PitchBook Founded 2022, formerly Hello, San Francisco, about $10.9 million raised, a 3 December 2025 venture deal, and an out-of-business status dated 16 April 2026.
  • Phind | Y Combinator Batch listing. Later coverage describes the page as marking the company inactive and pointing Royzen at a subsequent YC company. Confirm the live page before quoting a status line.
  • Phind AI shut down | Fello AI, 2026 A secondary reconstruction of the 12 January email, the 16 January stop date, refunds, and the 30 January deletion. Useful where it quotes the email. Its alternatives section is advertising.
  • Phind status notes | AnyGen, accessed 2026 Collects the Y Combinator inactive flag and the January dates. Treat traffic and motive claims in roundup sites as unverified.
  • Latent Space episode announcement | swyx, 3 November 2023 The host's summary of the fine-tune story, including a 74.7 percent HumanEval figure that does not match the model card's 73.8. Cited so the two numbers are not silently merged.

Additional Viewing

Load video
YouTube

"Beating GPT-4 with Open Source Models" (Latent Space, Michael Royzen, 3 November 2023, about 1 hour 18 minutes) is the conversation that belongs with the model section. Royzen explains how a small team fine-tuned Code Llama into a coding model that, on the benchmarks they cared about, sat next to GPT-4, and how Phind used retrieval so the answer could point at a real page. The host's numbers and the model card's numbers differ slightly. Listen for the method, then cite the card. Royzen on Phind's code model

Load video
YouTube

"Is Phind AI's Code Llama Fine-Tune Better Than GPT-4 Code Interpreter?!" (Ai Flux) is a 2023-era comparison aimed at the same claim as the model card: whether the open fine-tune was actually better at coding tasks than OpenAI's code tool of that moment. It is a demonstration, not an independent paper, and both products it compares have since been replaced. Use it to see what "better than GPT-4 at code" looked like in a screen recording. Phind's Code Llama versus GPT-4 code tools

Load video
YouTube

"Is Phind AI Search the New ChatGPT for Devs?" (SavvyNik) walks the consumer search page: a developer question, a generated answer, and the sources underneath. That is the product described in the first sections, recorded while it was still a plausible daily tool. Interface details in the video are not a current manual. Phind as a developer search engine

Load video
YouTube

"I Tried Every AI Coding Assistant" (Conner Ardman) was already embedded here. The title is a roundup of the 2023 assistant field (completion plugins, chat, and editor tools), which is the neighborhood Phind was trying to sit in, not a tour of Phind itself. A similarly titled video from the same creator is under a minute and barely touches this product, so do not treat the embed as a Phind review until you have watched this specific upload. Coding-assistant roundup, 2023

Retrieved from "https://primo.ai/w/index.php?title=Phind&oldid=54735"