SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Kimi K3 AI Model Can Now Run Locally on 29 GB RAM at 0.50 Tokens per Second

0
·1 views

A project shared on Hacker News demonstrates running the Kimi K3 language model locally using just 29 GB of RAM. The implementation achieves a processing speed of 0.50 tokens per second, making it accessible on consumer-grade hardware. The project, hosted on GitHub under sqliteai, has garnered 24 points and 4 comments on Hacker News. This development is notable as it lowers the hardware barrier for running large language models without cloud infrastructure.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

AI Coding Tools Routinely Hardcode Live API Keys, Posing Serious Security Risk

AI-powered code editors like Cursor frequently embed real API keys, passwords, and JWT secrets directly into source code, replicating a pattern common in tutorial and example code used for their training. Once a secret is committed to a Git repository, it becomes permanently embedded in the project's history, meaning deleting the line in a later commit does not eliminate the exposure. The underlying issue stems from AI models having no ability to distinguish between safe placeholder strings and live credentials, so they default to inline values that produce immediately runnable code. Developers are advised to load all secrets from environment variables and add a secrets-scanning tool such as Gitleaks as a pre-commit hook to prevent credentials from ever entering tracked files. If a live key is accidentally committed, it must be revoked and rotated immediately, as the history cannot be considered safe even after the line is removed.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How Bitcoin Transaction Fees Work: Size, Fee Rates, and UTXO Model Explained

Bitcoin transaction fees differ fundamentally from account-based blockchains because they are calculated using the UTXO model, where fees are based on transaction data size measured in virtual bytes rather than the amount transferred. The fee is determined implicitly by the difference between total input value and total output value, with leftover satoshis going to the miner. Developers building on Bitcoin must explicitly set a fee rate in sat/vB when constructing raw transactions. Best practices include dynamically querying mempool APIs for current fee estimates, implementing Replace-By-Fee signaling to allow fee bumping on stuck transactions, and avoiding dust outputs that fall below the network's minimum relay threshold.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds Search Engine From Scratch, Then Compares It to Apache Lucene

A software developer set out to understand how large-scale search bars like Google's actually work, studying techniques such as inverted indexing, BM25 ranking, Trie structures, and BK-trees. The exploration led to the discovery of Apache Lucene, an open-source search library that underpins many modern search systems, including MongoDB's Atlas Search. To compare approaches, the developer built two parallel projects: one coded from scratch and another leveraging Atlas Search on top of Lucene. A web crawler was configured to index Microsoft's .NET documentation from the web as the shared data source for both implementations. The experiment aimed to measure real differences in development time, lines of code, and search quality between reinventing the wheel and using an established library.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Engineer explains real differences between data migration, conversion, and ETL with project examples

A software engineer has outlined the practical distinctions between three commonly conflated data concepts: migration (a one-time system move), conversion (reshaping data during a migration), and ETL (a continuous pipeline between live systems). Drawing on real professional projects, the author describes a Sybase-to-Oracle consolidation where CSV exports, loose VARCHAR2 dump tables, and PL/SQL scripts were used instead of heavyweight tools like Talend or MuleSoft. The dump-table approach was chosen deliberately so that bad source rows would surface during transformation rather than crashing the load boundary. A third project involved building a MuleSoft-to-Salesforce integration as a portfolio piece, illustrating a genuine ETL use case where both systems remain active. The author's central takeaway is that teams over-engineer migrations by applying ETL tooling permanently when a single-run script would suffice.

0 comments Read more at DEV Community
Kimi K3 AI Model Can Now Run Locally on 29 GB RAM at 0.50 Tokens per Second · ShortSingh