1M-token needle tests on GB300: DeepSeek V4.1 Flash nails single hops, fails two-hop chains
Single-needle retrieval hit 16 of 16, but at 125K with thinking off DeepSeek scored 0 of 12 on two-hop questions, recovering to 12 of 12 when split or with thinking on; GLM took 48.8s to first token at 1M.
What breaks first in long context, from my 1M token needle tests on one GB300: Finding one fact is mostly solved. DeepSeek V4.1 Flash found 16 of 16 single needles at every length. Linking two facts isn't. Every miss, on every model, was a two-hop chain. At 125K with thinking off, DeepSeek went 0 of 12 on two-hop questions. Split the same questions into single hops: 12 of 12. Turn thinking on: 12 of 12. Waiting is the tax. GLM's first token at 1M took 48.8 seconds. Context length is a storage spec. Reasoning across it is a separate skill.