webAI releases open 8B model TwIL-8B on Hugging Face
Hesamation · x · 2026-11-05
webAI has published TwIL-8B model weights on Hugging Face. Its chat template reveals support for tool calling, configurable reasoning effort (including a low-effort mode), and thinking-history truncation, built on ChatML-style special tokens (<|imstart|>/<|imend|>). The post is just a link drop — no benchmarks or hands-on results yet.
More from Models
- Ripple claims frontier-level open-weight model trained in 32 minutes at one-tenth cost — annadgoldie · 2026-11-05
- Variable names matter: safety-flavored naming can push LLMs into over-cautious mode — jessi_cata · 2026-11-05
- Developers report Grok API instability: 13.58% error rate and 13.6s p90 latency in dashboards — ChrisUniverse · 2026-11-05
- Ramp's post-trained Qwen beats Opus on finance QA; 10,000-agent swarm finds 7 critical bugs — AI Engineer · 2026-11-05
- Vals AI: four of seven frontier models keep 80%+ scores at lowest reasoning effort for half the cost — ZacharyHuang12 · 2026-11-05
- Sailing sim race shows Chinese models lag frontier: "cheap tokens ≠ cheap agents" — shekitup · 2026-11-05