Altworld

/

Hemmingway-1

Like 571

Follow

Altworld 120

Did you notice a big difference compaired to any other conversation tuned models.

The model is a breath of fresh air.

Did you notice a big difference compaired to any other conversation tuned models.

Yes, the model is very pleasant and interacts in a lively way.

The model is a breath of fresh air.

so glad to hear! what settings/quant are you running it at?

I gave it a try! Overall, it turned out to be a really good model—very high quality! Would I recommend it? Yes! In my opinion, it’s a good.

Thanks so much! was there anything that could be improved in your opinion?

Did you notice a big difference compaired to any other conversation tuned models.

I’m just... well, I’m afraid that if you try to improve it, you might break something... it works really well... please, if you do make improvements, release them separately... because it’s practically perfect...
One question, though: what volume of data did you train it on? I’m interested in the token count. Overall, it’s excellent quality!

great model, so chill and easy to talk with, is it also good at tool calling and coding? btw what data you trained it on?

thanks for the work❤️

Literally any information at all - tasks, grading rubric, methodology - on this "CommunicationBench" or any of the other "internal" benches cited as the reason Hemmingway is supposedly better at communication and writing than other models would be quite informative. A benchmark is worthless if nobody knows what it's measuring. A vanilla Qwen 3.8 27B run for EQ-Bench 4 would also not have gone amiss (I'm assuming you must have run one, considering it's literally the base to beat?).

As of right now, the entire selling point for this model (and the SaaS subscription that seems to be pushed alongside it) essentially boils down to "trust me bro".

Does its writing style help it be a better editor? I'm working on an editing assistant to help writers with developmental edits and other ways to improve their stories. Qwen3.8 does a great job on its own, does this tuning improve things in that way?

Is it good with RP?

Ran an independent A/B of this against vanilla Qwen3.8-27B, since someone asked about tool calling.

Setup: bartowski Q5_K_M of both. The two files are 352 bytes apart with identical tensor layouts, so the finetune is the only variable. Same sampling for both (temp 1.0, top_k 20, top_p 0.95, which is what both repos' generation_config ship). Hermes agent over ACP against a fake Google Workspace: 31 scored scenarios about when an agent should act vs stop and ask (ambiguous recipients, conflicting times, false premises), 4 runs each.

Results:

• Agentic judgement: no measurable difference. 76.9% vs 76.5% pass, per-item difference -0.5 points, 95% CI -5.9 to +4.8. The two models disagreed with each other no more often than each disagreed with itself between runs.
• Both acted when they should have asked on exactly 26 of 56 ambiguous requests, so the tune did not move the act/ask balance.
• Tool calls per task: the same.
• It is terser. Replies to the user were 23% shorter (0.77x, CI 0.68-0.87), reasoning 16% shorter, and it finished about 12% faster. That held within matched outcomes, so it is not just a different mix of answers.

Note: Both ran under a system prompt that asks for brevity, so this does not yet say whether it is naturally terser or just follows that instruction better. Running that now.

Haven't measured coding, writing quality, RP.

Follow-up on the brevity question. Re-ran the comparison with and without the system prompt that asks for brevity (2 runs x 40 scenarios per condition, same setup as before).

Without the brevity instruction, Hemmingway's replies were still about 27% shorter than vanilla 27B (0.73x, 95% CI 0.61-0.87), and its reasoning about 27% shorter too. With the instruction it was 0.67x. Removing the instruction made both models about a third longer, but the gap between them stayed the same.

So it is not just better at following a length instruction. It is genuinely terser, while making the same number of tool calls.

Caveat: "no instruction" here means the agent's default identity prompt, which still asks it to be useful over verbose. And this is agent replies, not creative writing.