Skip to content
Notifications
Clear all

Has anyone benchmarked Continue's local model (Codestral?) against GPT-4 for code review?

2 Posts
2 Users
0 Reactions
3 Views
(@cloud_bill_shock)
Estimable Member
Joined: 2 months ago
Posts: 114
Topic starter   [#21253]

Everyone's pushing local models for "privacy" and "cost." But has anyone actually measured the cost of being wrong?

If Codestral misses a critical security flaw or a performance anti-pattern that GPT-4 catches, what's the real price? You're trading a direct LLM API cost for future tech debt and incident response.

* What's the false-negative rate?
* Has anyone compared feedback quality on a real, messy codebase?
* Are the suggestions actually actionable, or just generic?

The cloud bill for a bug in production can dwarf your entire year's AI spending.


show me the bill


   
Quote
(@ide_tinkerer)
Estimable Member
Joined: 3 months ago
Posts: 104
 

Exactly, that's the real hidden cost - the price of a missed warning. I've been testing Codestral locally via Continue for a few weeks, and while it's decent for catching simple syntax issues or style violations, it's surprisingly quiet on deeper architectural implications.

For example, it passed over a potential SQL injection in a dynamic query built with string concatenation in a code review last Tuesday, while the GPT-4 review I ran in parallel flagged it immediately with a specific parameterized query suggestion. That's the kind of "false negative" you can't afford.

The actionable advice point is huge too. Codestral's feedback often stops at "consider using a more efficient data structure," but GPT-4 tends to add "...like switching this list lookup to a dictionary, which would reduce this loop from O(n) to O(1)." The specificity changes how quickly you can act on it.


editor is my home


   
ReplyQuote