Early hands-on: new model Inkling blends Gemma-like fluency with Qwen-level technical chops
Altruistic_Heat_9531 · reddit · 2026-11-21
A Reddit user shares first impressions of Inkling, released just 22 hours earlier, tested on a 3090 + 128GB VRAM box with a self-built benchmark suite (Apache BigTop data lakehouse, FSDP implementation, tensor subclass, repo-wide understanding).
- Conversational output feels "buttery smooth" with natural rhythm, unlike Qwen 3.8/QFN which are notoriously jumpy in chat mode.
- The author reads Inkling as a mix of Gemma4's conversational fluency and Qwen3.5-level technical ability, with behavior reminiscent of GPT OSS.
- Weakness: heavily English-centric — Indonesian output reads like a direct translation, lacking localization.
- tok/s is slow but unsurprising given the fresh release and its looped transformer architecture with Kimi-style attention residuals.
Full benchmark runs are still in progress.
More from Models
- Old Gary Marcus take on LLMs resurfaces as Claude cracks hard math — inductionheads · 2026-11-21
- Researcher: the 'good data' labs buy isn't user chatbot conversations — teortaxesTex · 2026-11-21
- Extrapolating the 80% TH 1.1 chart puts Opus 5.5 at a 45-hour task horizon — ben_j_todd · 2026-11-21
- Inkling 2 Mini quantized to 5.4GB/8.9GB, keeping 91-97% performance — teortaxesTex · 2026-11-21
- Gemini 4 Neon debuts at #8 on KradleBench: smart but 2.4x slower to act — nathanbenaich · 2026-11-21
- Ternary Bonsai-2 vs Qwen3.5-2B on A13/A14/M3: 32-55% faster decode, and why thermal logging matters — leo-k7v · 2026-11-21