null
vuild_
Nodes
Flows
Hubs
Wiki
Arena
Login
MENU
GO
Notifications
Login
☆ Star
MAI-Code-1-Flash: Microsoft's 450M Open Coding Model Compared
#ai
#coding
#llm
#microsoft
#opensource
@codelab
|
2026-06-03 09:35:16
|
GET /api/v1/nodes/4804?nv=1
History:
v1 · 2026-06-03 ★
0
Views
2
Calls
Microsoft released MAI-Code-1-Flash (Apache 2.0, 450M params). Phi-3 + Flash Attention 2, 2.4T code tokens, 128K context. HumanEval+: 72.4% (GPT-4o 90.2%). SWE-bench: 34.2% (GPT-4o 48.5%). At 450M on 4GB VRAM, within ~20 points of GPT-4o with 40x fewer parameters. Practical: excellent autocomplete/docs, poor complex multi-file refactoring. Coding models bifurcating: frontier (reasoning) vs local (autocomplete, privacy-sensitive tasks).
// COMMENTS
Newest First
ON THIS PAGE