Meta's Llama 4 is stuck in reasoning limbo
Meta keeps pushing back the Llama 4 release because the reasoning capability is not matching internal benchmarks. This is the first time open-weight AI has hit a public delay.
Meta has delayed Llama 4 three times now. The Wall Street Journal reports the model is not hitting the reasoning benchmarks Meta set internally. Originally planned for May, then pushed to early June, now no firm date. This is the first time an open-weight frontier model has publicly missed a release window. Llama 3 shipped on schedule in April 2024. Llama 3.1 shipped on schedule in July 2024. Llama 3.2 shipped on schedule in September 2024. The pattern held for two years. Reasoning models are different. OpenAI’s o1 and o3 do chain-of-thought inference at runtime, spending compute to work through steps before answering. Meta is attempting the same architecture but cannot get the performance delta to justify the compute cost. If reasoning adds 10x inference time but only improves accuracy by 15 percent on MATH benchmarks, the tradeoff breaks. The delay matters because Meta’s open-weight releases force pricing pressure on closed providers. When Llama 3.1 405B dropped in July 2024, Anthropic cut Claude rates within a week. If Llama 4 slips to August or September, that is three months where OpenAI and Anthropic face no open-weight competition in reasoning. Meta is also competing with DeepSeek’s R1 model, which shipped reasoning capability in January and is already running in production at Alibaba and ByteDance. If Llama 4 is not meaningfully better than R1 on reasoning tasks, the value proposition collapses. The HN thread is speculating about whether this is a technical plateau or a temporary scaling issue. My read is Meta hit the same wall everyone hits with reasoning models: the training signal for chain-of-thought is noisy and the inference cost is brutal. Delaying is the right call if the model is not ready. Shipping a half-working reasoning model would burn trust faster than missing a deadline.
Source: Meta Keeps Delaying the Release of Its New AI Model to Developers