webAI Air-26B-A4B Card - wIL Track B - MuSR: leaderboard
MuSR multistep soft reasoning over long narratives (murder mysteries, object placement, team allocation), run in webAI's wIL harness Track B with thinking off, greedy decoding and up to 300 examples, through vLLM with Air's quantized weights expanded to BF16 format. Accuracy on a 0-1 scale. From webAI's Air-26B-A4B model card (November 4, 2026): Google's BF16 Gemma 4 26B A4B (IT) against webAI's 4-bit Air quantization of it, both run by webAI on the same items with the same settings, one sampled run each (Air with an 8-bit KV cache, BF16 with a 16-bit one). Higher is better.
Metric: Accuracy (0-1). Source: huggingface.co. Saturation forecast: Around February 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 26B A4B (IT) (Non-reasoning) | 0.64 | #545 (Gemma 4 26B A4B (IT)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=webai-air-26b-a4b-card-wil-track-b-musr · How It Works · Data refreshed daily, snapshot 2026-11-12.