Local AI Advisor

Can Mac Studio M5 Ultra 256 GB run Kolibri-1 (78B-A3.5B)?

256 GB unified · 1228 GB/s memory bandwidth · Kolibri-1 (78B-A3.5B) has 78.1B parameters (3.46B active per token).

Short answer

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

Best quantization for chat: FP8 (73 GB of weights) on LM Studio. Generation: 100–170 tok/s, leaving about 172 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 WebUIExcellentFP8100–170 tok/sunder 1 s8KDetails
Coding questionswith ContinueExcellentFP8100–170 tok/sunder 1 s16KDetails
Coding in a repositorywith AiderAcceptableFP895–160 tok/s≈2 s32KDetails
Agentic codingwith OpenCodeAcceptableFP890–150 tok/s≈2 s32KDetails
Long documents (64K)with Open WebUIExcellentFP880–130 tok/s≈5 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.