Local AI Advisor

Can Desktop PC · GeForce RTX 4070 Ti SUPER · 32 GB RAM run Recursive Flint (40B-A2.5B)?

16 GB VRAM + 32 GB RAM · 672 GB/s memory bandwidth · Recursive Flint (40B-A2.5B) has 39.8B parameters (2.5B active per token).

Short answer

It loads, but it's borderline even for chat.

Best quantization for chat: MXFP4 (20 GB of weights) on Ollama. Generation: 50–80 tok/s, leaving about 0.0 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 WebUIBorderlineMXFP450–80 tok/sunder 1 s8KDetails
Coding questionswith ContinueBorderlineMXFP445–75 tok/s≈3 s16KDetails
Coding in a repositorywith AiderBorderlineMXFP435–60 tok/s≈10 s32KDetails
Agentic codingwith OpenCodeBorderlineMXFP435–55 tok/s≈15 s32KDetails
Long documents (64K)with Open WebUIBorderlineMXFP420–35 tok/s≈45 s64KDetails

Speeds are bandwidth-based estimates; no direct benchmark exists for this pair. “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.