Sub-32B Open Weights Hit GPT-5 Intelligence: Qwen3.5 27B Scores 42, Gemma 4 31B Scores 39
The sub-32B weight class cleared a threshold this week. Alibaba’s Qwen3.5 27B (Reasoning) and Google DeepMind’s Gemma 4 31B (Reasoning) both land in GPT-5 territory on the Artificial Analysis Intelligence Index — 42 and 39 respectively, matching GPT-5 (medium) and GPT-5 (low).
That comparison comes with a sharp asterisk.
Where the Gap Closed
On agentic and reasoning tasks, the open models are genuinely competitive:
- Qwen3.5 27B Agentic Index: 55 — ahead of GPT-5 (medium) at 46
- Gemma 4 31B on TerminalBench Hard: 36% vs GPT-5 (low) at 27%
- Gemma 4 31B on HLE: 23% vs GPT-5 (low) at 18%
Both families include reasoning and non-reasoning variants, multiple sizes, and native multimodal input — feature parity with proprietary mid-tier models.
Where the Gap Didn’t Close
Factual knowledge and hallucination avoidance remain frontier-gated. AA-Omniscience scores tell the story:
| Model | AA-Omniscience |
|---|---|
| Qwen3.5 27B | -42 |
| Gemma 4 31B | -45 |
| GPT-5 (medium) | -10 |
| GPT-5 (low) | -10 |
A 32-35 point gap on factual recall means these models code well and reason structurally, but still confabulate at rates that matter in production deployments that rely on accurate knowledge retrieval.
What It Means
Open-weights intelligence is no longer a lagging indicator of proprietary capability — it’s a concurrent track. Sub-32B models that would have been mid-tier one year ago now clear the GPT-5 floor on task performance benchmarks.
The practical consequence: for agentic workflows where factual precision is bounded (code generation, structured reasoning, tool use), Qwen3.5 27B and Gemma 4 31B are now viable cost-zero alternatives to mid-range API calls. Both are Apache 2.0 licensed and run on consumer hardware.
The remaining moat for frontier proprietary models is knowledge depth, not reasoning architecture.