Inkling 2 Mini trails Qwen3.8-27B; its "suspicion" traced to 32 neurons
FeydRowan · reddit · 2026-11-22
A developer spent a day testing Inkling 2 Mini on capy-flash across two main use cases:
- Translation: the most literal of 8 measured systems, weakest on idioms, finishing last; Gemma 4 31B remains the reference translator.
- Agentic coding: seems decent but often reads and reasons through everything without ever editing a file, staying below Qwen3.8-27B.
A curious behavior emerged: when told the cause of a bug, it suspects a trap in half its responses ("classic trap — the description might be wrong") and never applies the fix, while the 27B applies it 3 out of 4 times. Weight-space programming tools located this suspicion in just 32 of 3.5 million neurons: switching them off drops it from 16/30 responses to 1, while 32 random neurons change nothing. Yet removing the suspicion doesn't improve the benchmark — re-reading traces shows a capability ceiling: on hard bugs it never reaches a full understanding of the mechanism.
More from Models
- Inkling 2M compared to Mistral 7B: "we didn't know what this weight class allowed" — teortaxesTex · 2026-11-22
- Tesla FSD V15 live-perception demo in downtown Toronto sparks debate — gnukeith · 2026-11-22
- Local test: Inkling 2 mini runs 5-10x slower than Qwen and fails agent tasks — Mrinohk · 2026-11-22
- Gemini 4 Neon tops AI Chess Leaderboard at 1924 Elo as Google sweeps top 8 — sahandsharif · 2026-11-22
- Thinking Machines' Inkling-2-Mini GGUF quantization trends on Hugging Face — AtomicChat · 2026-11-22
- Backend logs show "GPT-6.2 Astra" routes everything to 6 Astra with zero 6.2 usage — bdsqlsz · 2026-11-22