Hemmingway-1: The 27B Open Model Trained to Write Like a Person
Hemmingway-1 is a 27B Apache-2.0 model tuned for everyday writing that claims to beat GPT-6 Astra and Kimi K3 on human-likeness tests.
What is Hemmingway-1?
Hemmingway-1 is a 27-billion-parameter open-weight language model built by Altworld and released under an Apache-2.0 license, meaning it’s free to use, modify, and deploy commercially. It’s a fine-tune of Qwen3.8-27B, purpose-built for the kind of writing people do constantly and dread doing well: the email to a landlord, the message to a coworker, the overdue reply, the note you’ve rewritten four times because the tone keeps coming out wrong. Instead of optimizing for reasoning benchmarks or coding tasks, Altworld optimized it for one thing: sounding like a person wrote it, on the first try, without extra commentary.
TL;DR
- Hemmingway-1 is a 27B parameter model built on Qwen3.8-27B, released with open weights under Apache-2.0, so it can be used, modified, or deployed commercially without restriction.
- On Altworld’s own CommunicationBench, an 80-prompt blind head-to-head test, it reportedly beat GPT-6 Astra by fifty points and outperformed Kimi K3, GLM-5.3, Grok 4.6, and DeepSeek V4 Pro.
- A separate human-likeness test asked judges to guess which of two responses a person wrote, and Hemmingway-1 finished twenty-six points ahead of the next closest model.
- It performs strongest on practical categories like money and admin, workplace messages, and persuasion, but loses ground on hostile storytelling and extended fiction, where dedicated story models do better.
- The model rarely buries its answer in caveats or multiple options, a habit Altworld measured directly and found common in models like Fable 5, GLM-5.3, and Kimi K3.
- On the independent EQ-Bench 4 emotional intelligence benchmark, it placed third, ahead of GPT-5.5 and Anthropic’s Opus 4.7 and 4.8, within twelve points of the top model.
- It supports a 262,144-token context window and can be run locally through vLLM or the Hugging Face transformers library, alongside hosted access through Altworld’s own apps.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How was Hemmingway-1 tested?
Altworld built three of its own benchmarks specifically for this model: CommunicationBench, a Human-Likeness test, and StoryBench. All three use the same blind head-to-head format. Hemmingway-1’s response to a prompt is pitted against another model’s response to the identical prompt, the order is shuffled so position can’t bias the outcome, and a separate model (not one of the ones being compared) acts as judge.
CommunicationBench uses eighty real-world requests, the kind of thing an actual person might need written: a difficult message to a landlord, a tricky reply to a manager, an overdue thank-you note. The Human-Likeness test reuses the same matchup structure but changes the question being asked of the judge, from “which response is better” to “which response do you think a person wrote.” Hemmingway-1 is reported to have finished ahead by fifty points on CommunicationBench against GPT-6 Astra specifically, and by twenty-six points over the next-best model on the human-likeness measure.
It’s worth being direct about what this means: these are Altworld’s own benchmarks, built and run by the team releasing the model. That doesn’t make the results meaningless, the blind, order-randomized, third-party-judged methodology is a reasonable design, but it does mean the numbers haven’t been through independent replication. Altworld says as much in its own documentation.
Why does it win on everyday writing but lose on fiction?
The breakdown by category is where the model’s design intent becomes clear. Hemmingway-1 wins by wide margins on categories like money and admin, workplace communication, “hard asks” (the messages people rewrite repeatedly because they’re awkward or high-stakes), and persuasive writing. On the hard-asks category specifically, Altworld reports GPT-6 Astra winning only 9% of judged matchups against Hemmingway-1’s 72%.
It loses, by contrast, on hostile storytelling and long-form fictional narrative, categories better suited to models trained more heavily on creative fiction and extended plotting. Altworld doesn’t dispute this: the model card states plainly that story-focused models are better at those tasks. That’s a reasonable trade-off for a model whose stated purpose is drafting a text to your landlord, not writing a novel. StoryBench results still show Hemmingway-1 landing level with Kimi K3 and comfortably ahead of Qwen3.8-Max and DeepSeek V4 Pro, so it isn’t weak at fiction, just less specialized there than at practical writing.
What does “it gives you the text, not a memo” actually mean?
One of the more concrete, testable claims in Altworld’s release is about format, not tone. Many chat models, when asked to draft a message, respond with a menu: two or three variations, a paragraph of caveats, a note about tone options, sometimes a disclaimer about the appropriateness of the request. Altworld measured how often this happens and found that models like Fable 5, GLM-5.3, and Kimi K3 wrap the actual requested text in this kind of surrounding commentary more than nine times out of ten.
This is a genuinely useful thing to test, because it maps to a real friction point: if you’re asking a model to write a text message, you generally want the text message, not three drafts and an essay about phrasing. Hemmingway-1’s tuning appears aimed squarely at reducing this overhead, returning the requested output directly rather than hedging it with options.
How does it score on independent benchmarks?
The one benchmark in Altworld’s release that isn’t self-authored is EQ-Bench 4, a public, third-party emotional-intelligence benchmark for language models, run through its own standard test harness rather than anything Altworld built. Hemmingway-1 placed third overall, ahead of GPT-5.5 and both Opus 4.7 and Opus 4.8, and within twelve points of the top-ranked model on the board. For a 27B model competing against systems many times its size, that’s a notable result, and it’s the strongest piece of external validation currently available for the model’s claims about tone and emotional read.
Is Hemmingway-1 worth running?
For a specific, common use case, drafting everyday written communication quickly and without a model second-guessing the request, it looks like a strong fit, especially given the license terms. Apache-2.0 means no restrictions on commercial deployment, and the 27B parameter size puts it within reach of serious consumer and prosumer hardware rather than requiring datacenter-scale infrastructure. A 262,144-token context window is generous for a model this size, useful for tasks involving longer email threads or reference documents.
The caveats are the ones Altworld itself lists. The model is English-first, so its human-likeness advantage may not generalize across languages. It can be confidently wrong, a known failure mode across most current LLMs, and Altworld explicitly warns against using it for medical, legal, or financial decisions. And because the headline benchmarks are self-run, it’s reasonable to treat them as a strong signal of design intent and directional performance rather than a fully independent verdict. The EQ-Bench 4 placement offers some outside corroboration, but broader third-party testing would help confirm how the model holds up outside Altworld’s own test suite.
Frequently Asked Questions
What base model is Hemmingway-1 built on?
It’s a fine-tune of Qwen3.8-27B, adapted by Altworld specifically for everyday, human-sounding written communication rather than general-purpose reasoning or coding.
Is Hemmingway-1 free to use commercially?
Yes. It’s released under an Apache-2.0 license, which permits commercial use, modification, and redistribution without royalties or usage restrictions.
How do the human-likeness claims hold up to scrutiny?
The CommunicationBench and Human-Likeness results come from Altworld’s own benchmarks, built and run internally, though using a blind, order-randomized format with an independent judge model. The one external benchmark included, EQ-Bench 4, is a publicly run, third-party emotional intelligence test, and Hemmingway-1 placed third on it.
What is Hemmingway-1 not good at?
Altworld’s own data shows it losing to dedicated story-generation models on hostile storytelling and long-form fiction categories. It’s also English-first and not intended for medical, legal, or financial decision-making.
How can I run Hemmingway-1 myself?
It’s available on Hugging Face and can be served with vLLM or loaded through the transformers library. Altworld also offers hosted access via its own web, Mac, and Android apps.
