Local AI Advisor

Can Mac Studio M3 Ultra 96 GB run Recursive Flint (40B-A2.5B)?

96 GB unified · 819 GB/s memory bandwidth · Recursive Flint (40B-A2.5B) has 39.8B parameters (2.5B active per token).

Short answer

Yes, excellent for chat and comfortable or better for every workload we test.

Best quantization for chat: MXFP4 (20 GB of weights) on LM Studio. Generation: 180–240 tok/s, leaving about 65 GB free for your other apps.

How it rates for each workload

Same model and machine, different jobs. Agents and long documents send far bigger prompts than chat, so they need much more speed and memory.

WorkloadVerdictQuantGenerationFirst replyContext
Chatwith Open WebUIExcellentMXFP4180–240 tok/sunder 1 s8KDetails
Coding questionswith ContinueExcellentMXFP4160–220 tok/s≈2 s16KDetails
Coding in a repositorywith AiderComfortableMXFP4140–180 tok/s≈4 s32KDetails
Agentic codingwith OpenCodeComfortableMXFP4130–170 tok/s≈6 s32KDetails
Long documents (64K)with Open WebUIExcellentMXFP485–120 tok/s≈15 s64KDetails

Speeds are estimates calibrated against measured benchmarks on similar hardware. “First reply” is the time to read the workload’s opening prompt. How we calculate

Different tool, context size or workload?

These pages use typical settings. The full check models your exact setup.