Altworld's Hemmingway-1 Isn't Apache-Licensed, Despite Reports
A.I. / news
Altworld's Hemmingway-1 Isn't Apache-Licensed, Despite Reports
Hugging Face's own metadata says the 27-billion-parameter writing model is noncommercial only, contradicting at least one widely read AI blog.
Altworld published Hemmingway-1, a 27-billion-parameter open-weight model tuned to write everyday messages, under a noncommercial license, contradicting an independent AI blog's report that described the release as Apache 2.0.
Hugging Face's own metadata for the repository lists the license field as cc-by-nc-4.0, and the model card states plainly: "Hemmingway-1 is free for personal, research and other non-commercial use under CC BY-NC 4.0. Download it, run it, fine-tune it and share it, as long as you credit Hemmingway-1. Commercial use needs a separate agreement with us," naming a contact address for that agreement. The repository's metadata shows the weights were last updated Sept. 22.
Where the two accounts diverge
MindStudio, an independent AI blog that covered the release, described Hemmingway-1 as carrying an "Apache-2.0 license, meaning it's free to use, modify, and deploy commercially." That is not what Altworld's own model card says.
| Source | License stated |
|---|---|
| Hugging Face model card (Altworld) | CC BY-NC 4.0, noncommercial only |
| MindStudio | Apache-2.0, "free to... deploy commercially" |
Hemmingway-1 is a fine-tune of Qwen3.8-27B, the dense model Alibaba's Tongyi Lab released Aug. 14 under a genuinely permissive Apache 2.0 license. Nothing in that base license obligates Altworld to keep its fine-tune open to commercial use, and licensing a derivative more restrictively than its base model is common and allowed. The dispute here is not about what Altworld is permitted to do. It is about what outside coverage told readers they were getting.
What Altworld claims, and who checked it
Altworld says the model beats OpenAI's GPT-6 Astra by 50 points on its own CommunicationBench, and scores 26 points ahead of the next-closest model on a human-likeness test it also built itself. On what it calls "hard asks," Altworld says Hemmingway-1 was preferred 72 percent of the time against 9 percent for GPT-6 Astra. Those are the company's own self-run comparisons.
One outside figure exists: MindStudio reported that Hemmingway-1 placed third on EQ-Bench 4, a public emotional-intelligence leaderboard, ahead of GPT-5.5 and Anthropic's Opus 4.7 and Opus 4.8, and within 12 points of the top-ranked model. MindStudio also flagged that Altworld's other benchmarks are self-run and self-authored, without independent replication, and warned the model "can be confidently wrong" and should not be used for medical, legal or financial decisions.
Altworld's own testing puts Hemmingway-1 level with Kimi K3 on what it calls StoryBench, and 504 points above the Qwen3.8-27B base model it was fine-tuned from, a category MindStudio said is not where the model is strongest: dedicated story-writing models, it said, still outperform Hemmingway-1 on long-form fiction, and its non-English performance has not been shown to match its English results. Altworld's model card lists a 262,144-token context window and gives a single contact for commercial licensing, an email address for a person identified only as Luka.
The pattern echoes a licensing gap this site found in Edge0's Audio8 speech model earlier in September, where a model built partly on a Qwen decoder carried a non-transferable non-commercial restriction underneath an Apache 2.0 badge the publisher had not corrected. Altworld's own listing is accurate. The confusion this time sits one level downstream, in how a widely read blog described it.
Sources
More in A.I.
- 01Anthropic's Levent Alpöge Claims a Claude-Assisted Riemann Proof, and Lean Checks Are InImperial's Kevin Buzzard and MIT's Andrew Sutherland have examined the formalization, according to Scientific American. No journal review has been reported.
- 02Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 03Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 04Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.