Building for Real Conversation
- Can you do me a favor and not immediately go into ai helper mode for every tiny little thing like that?
- You literally sound like a toaster with no personality
That's what people text us when a reply falls flat. Frontier models are post-trained to excel at coding and agentic tasks, so it's easy to ask them for something. But it hasn't felt good to tell a model something.
We built Tomo's first model, Dachi 1.0, to change that.
Dachi 1.0 is built for everyday conversation, post-trained on a base of DeepSeek V4.1 Flash. The training process was informed by synthetic data pipelines and anonymized conversation analytics spanning nearly half a billion messages.
Head to head
We put Dachi 1.0 up against other models live in production, including those at the frontier. Users returned to Tomo more when served Dachi, across time windows from the next 10 minutes to 4 hours, and replied to Dachi more often even when it didn't ask a question.
Texted back within 10 minutes
vs. GPT-6.1 Sol- Dachi 1.0+8%
Claude Opus 5.5+2%
GPT-6.1 Sol0%
Muse Spark Open−2%
Texted again within 4 hours
vs. GPT-6.1 Sol- Dachi 1.0+6%
Claude Opus 5.5+4%
GPT-6.1 Sol0%
Muse Spark Open+1%
Answered when it asked a question
vs. GPT-6.1 Sol- Dachi 1.0+4%
Claude Opus 5.50%
GPT-6.1 Sol0%
Muse Spark Open−7%
Answered when it didn't ask a question
vs. GPT-6.1 Sol- Dachi 1.0+5%
Claude Opus 5.5+2%
GPT-6.1 Sol0%
Muse Spark Open+3%
Messages per user
vs. GPT-6.1 Sol- Dachi 1.0+22%
Claude Opus 5.5+4%
GPT-6.1 Sol0%
Muse Spark Open+1%
Time to first reply, median
Lower is better- Dachi 1.04.2s
Claude Opus 5.57.1s
GPT-6.1 Sol5.4s
Muse Spark Open4.2s
38% fewer complaints than its base model
Compared to its base model DeepSeek V4.1 Flash in production, we also noticed Dachi 1.0 had significantly fewer complaint analytics events. This was especially so for complaints related to tone and answer quality.
Complaints about tone
vs. DeepSeek V4.1 Flash- Dachi 1.0−38%
DeepSeek V4.1 FlashBaseline
Complaints about answer quality
vs. DeepSeek V4.1 Flash- Dachi 1.0−38%
DeepSeek V4.1 FlashBaseline
One text, three replies
Dachi 1.0 is tuned to be delightful to text. It reacts the way a person would, matches your energy and keeps the thread going. Frontier models fall back on the assistant script: a congratulations, a feelings check and an offer to help.
- 6:42Tomo
iMessage
Today 6:42 PMthey offered me the jobWAIT WHATthey offered it?? that's huge, you must be buzzing right nowok i need the whole story. are you taking it??Dachi 1.0In the Tomo harness - 6:42
Opus 5.5iMessage
Today 6:42 PMthey offered me the jobThat's big news! How are you feeling about it? Excited, nervous, still deciding?If you want to talk anything through, like the offer itself, negotiating, or whether to take it, I'm happy to help. Or if you just want to celebrate, congrats! 🎉Claude Opus 5.5In a harness optimized for Opus - 6:42
GPT-6.1 SoliMessage
Today 6:42 PMthey offered me the jobCongratulations! 🎉 How are you feeling about the offer—excited, unsure, or a bit of both?GPT-6.1 SolIn a harness optimized for Sol
Frontier-level, at a fraction of the cost
We maintain rigorous internal eval suites for tone, judgment, and temporal understanding. On TomoBench, Dachi 1.0 matches Claude Opus 5.5 at a fraction of the cost, and beats every other model under $2 per 1,000 replies.
- Dachi 1.0: 84.0% on TomoBench at $1.23 per 1,000 replies
- DeepSeek V4.1 Flash: 82.6% on TomoBench at $1.15 per 1,000 replies
- GPT-6.1 Sol: 87.1% on TomoBench at $10.05 per 1,000 replies
- Gemini 4 Argon: 84.2% on TomoBench at $29.31 per 1,000 replies
- Claude Opus 5.5: 84.2% on TomoBench at $46.10 per 1,000 replies
- GPT-6 Luna: 82.7% on TomoBench at $0.72 per 1,000 replies
- Muse Spark Open: 76.0% on TomoBench at $10.85 per 1,000 replies
Each model runs inside the full Tomo agent. TomoBench is made up of many pillars, including:
- Conversationality
- Whether the model texts like a person, asking the right number of questions, following the right thread, and keeping replies the right length.
- Judgment
- Whether the model makes the social and practical calls a good friend would, without rubber-stamping or overcorrecting.
- Temporal Understanding
- Whether the model reasons correctly about relative dates and schedules reminders for the right times.
- Subagent Routing
- Whether the model hands each task to the right one of its 12 subagents, and knows which tasks to handle itself.
No regressions in core capabilities
Tomo employs model routing to ensure the right conversation turns use the right models, but it's important Dachi maintains its ability to call tools, invoke subagents, and understand temporal information. Training to improve prose can easily regress agentic capabilities, and we don't find any evidence that Dachi has regressed on any of these in our benchmarks.
Subagent Routing and Tool Calling
- Dachi 1.090%
DeepSeek V4.1 Flash89%
GPT-6.1 Sol87%
Gemini 4 Argon90%
Claude Opus 5.587%
GPT-6 Luna79%
Muse Spark Open78%
Temporal Understanding
- Dachi 1.088%
DeepSeek V4.1 Flash87%
GPT-6.1 Sol97%
Gemini 4 Argon96%
Claude Opus 5.594%
GPT-6 Luna88%
Muse Spark Open91%
How we trained it
Dachi 1.0 is post-trained with DPO on a training set created with a custom sampling strategy, described below, which was manually verified by our team.
- 01Sample
We generate several replies to each conversation from the base model, and prompt some of them to vary slightly on certain axes, like length, warmth, or curiosity.
- 02Rank
We rank each set of replies, some by hand and the rest with a judge model tuned on our own picks.
- 03Pair
Better replies are paired against worse ones, without the prompt that varied them, so Dachi learns to give the better reply on its own.
We chose its base, DeepSeek V4.1 Flash, after comparing it with other open models, including GLM 5.3 Flash, Muse Spark Open, and DeepSeek V4.1 Flash 1017. For each one, we ran GEPA to build a custom harness. DeepSeek V4.1 Flash came out ahead in our testing, and it's still the base we'd choose as newer models have come out.
What's next
This work was part of an early effort to make Tomo better for our users, at economics that will scale with their growth. There is lots more to do, but the feedback we've gotten so far tells us we're on the right track.
Authors
Varun Nair, Raymond Chen, Justin Quan
Acknowledgements
We thank Fireworks AI for their support in training Dachi 1.0.
Say hi.
Tomo is for everyone betting on themselves every day. Dachi is the model that helps them keep showing up.

