Llama 4 Maverick

The user layer is dominated by one HackerNews thread about Cerebras' claimed world record — 2,500 tokens/s inference on the 400B-parameter Llama 4 Maverick — and most of it is skepticism about the hardware claim, not the model. Top commenters (+93) argue the record is single-query-only and ignores batched throughput ('people run potentially hundreds of queries in parallel'), that Cerebras extracts '<0.1%' of its theoretical memory bandwidth vs Nvidia's '80% or higher', that CSE-3 PEs carry only 48kB of RAM each, and that production is capped at roughly 250-300 wafers/year; one comment flags the CEO's past felony guilty plea for accounting fraud. Counter-commenters defend on-chip SRAM model storage ('20 [WSE-3s] are enough for some of the largest open models') as escaping the memory-loading bottleneck. Model-relevant corrections also appear: Maverick is NOT 'the largest and most powerful' Llama 4 — the unreleased Behemoth is. From YouTube commenters: an informed take clarifies Maverick is a Mixture-of-Experts model where only a fraction of parameters (~17B) are active per query, making '17B parameters running on a single GPU... a lot more reasonable than 100s' — while another argues Meta 'still hasn't figured out that the size of the model doesn't necessarily correlate with the quality of its output.' General mood is hype fatigue: 'AI doesn't actually need to be able to do a job, it just needs to be good enough to convince the C-level execs that it can.' Notably absent: any first-hand report of using Maverick for coding, writing, or reasoning tasks.

No aggregate ratings were found for this product during the last harvest.