TwIL-8B hands-on: near-frontier coding at 8B, but 5x slower than a 9B peer on RTX 3080

solyarisoftware · x · 2026-11-06

The newly open-sourced TwIL-8B (8.8B params) claims 93% of frontier performance on coding, 92% on instruction following, and 90% on contest math.

A third-party tester ran the verbose Q4KM build against the 9B NeoHorse 1 on a 10GB RTX 3080 with reasoning off:

Transferred well: tool calls 28/30 (vs 30/30); light briefing 9.0/10; daily organization PA1 scored 54.7, beating NeoHorse's 48. Turning reasoning on lifted PA1 to 79.
Didn't transfer: context search at 16K scored just 4.75/10 (vs 9.7) — a real capability gap.
The dealbreaker is speed: 16 tok/s with thinking off, 7 with it on, versus 89 tok/s for NeoHorse on the same card (5.4x faster). On PA1, average runtime balloons from 54 seconds to 21 minutes.

Verdict: impressive intelligence for 8B on short tasks, but poorly suited as an all-purpose agent brain.

More from Models