27B dense (26,895,998,464 BF16 parameters)
262,144 tokens
Apache 2.0
Open-weight, English-first writing and chat model post-trained from Qwen3.8-27B, specialized for everyday messages, emails, difficult interpersonal requests, dialogue, and creative writing that avoids sounding machine-generated or surrounding the requested text with commentary. Hemmingway-1 retains the dense 64-layer Qwen3.8-27B hybrid architecture, alternating three linear-attention layers with one full-attention layer for 48 linear and 16 global-attention layers. The official configuration uses a 5,120-dimensional hidden state, 17,408-dimensional feed-forward layers, 24 query heads, four key/value heads, a 248,320-token vocabulary, interleaved mRoPE with a 10,000,000 base, and a 262,144-token context window; a separate one-layer multi-token-prediction module is shipped alongside the twelve BF16 model shards. The included chat template supports thinking and non-thinking generation, with xhigh, medium, and low reasoning-effort settings, and includes structured tool-call formatting inherited from the Qwen family. Altworld reports that Hemmingway-1 won its internally developed CommunicationBench, consisting of eighty real writing requests evaluated through blind pairwise comparisons run in both orders, beating Fable 5.1 and finishing fifty points ahead of GPT-6 Astra. On the lab's Human-Likeness evaluation it reportedly finished twenty-six points ahead of the next model, while its category breakdown reports a 72% human-likeness preference rate on difficult asks versus 9% for GPT-6 Astra. Altworld also reports that the model commonly returns the requested message directly rather than wrapping it in options, preambles, or explanatory notes. On the externally developed EQ-Bench 4, the model card reports a third-place finish ahead of GPT-5.5, Opus 4.7, and Opus 4.8 and within twelve points of the leading model. On Altworld's internal StoryBench it reportedly performs level with Kimi K3 and 504 points above the Qwen3.8-27B starting checkpoint. The developer explicitly identifies hostile storytelling and long story turns as relative weaknesses and notes that the model is English-first and may sound confident when incorrect. The release provides standard Transformers loading instructions and an example vLLM server command using the full 262,144-token context.