Sub-32B open models now match GPT-5-tier index scores and fit on one H100, but trail badly on factual recall, so they suit agentic tasks more than knowledge lookups.
Key takeaways · AI-distilled
Artificial Analysis reports that Qwen3.5 27B (Reasoning) matches GPT-5 (medium) at 42 and Gemma 4 31B (Reasoning) matches GPT-5 (low) at 39 on its Intelligence Index. Both fit on a single 80GB H100 in BF16 and can run on a MacBook when quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition →.
The composition differs from GPT-5: both open models score far worse on AA-Omniscience (-42 and -45 vs -10), which Artificial Analysis ties to smaller parameter counts that reasoning effort alone cannot offset.
Their gains come mainly from agentic work and reasoning: Qwen3.5 27B scores 55 vs 46 for GPT-5 (medium) on the Agentic Index, and Gemma 4 31B beats GPT-5 (low) on TerminalBench Hard (36% vs 27%) and HLE (23% vs 18%).
The tradeoff between the two is tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →: Qwen3.5 27B used 98M output tokens to reach 42, while Gemma 4 31B used 39M for 39. The ~3-4B active-parameter tier now has six models from five labs, led by Qwen3.5 35B A3B at 37.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
