Dwight is Zime's open model, built for one job: reading sales calls. Its first answer about a call lands in 0.13 seconds and ten answers in 0.31, each with a confidence you can trust. It writes insights that quote the buyer word for word, and runs on one GPU you own.
True or false.
Never “it depends.”
Shown at 5× speed. Dwight on one H100 with vLLM. Jev and GLM 5 through their APIs. Times from the model card.
$205 a month where a frontier API costs $8,690, for 100,000 decisions and 100,000 insights.
0.02 s for each further question on the same call. GLM 5 takes 1.9 s for every one.
Every excerpt it returned was found word for word in the transcript.
ECE of 0.02 to 0.05 at scale. A confidence is a number you can set a threshold on.
General models are built to talk about anything. Dwight is fine-tuned on 500+ sales-call behaviors, so it spends every parameter on the call in front of it.
Yes or no, a choice, or a 0 to 10 score, as one token with a calibrated probability. Send the sure answers straight to your CRM and the close calls to a person.
Every insight carries the timestamp and the buyer's exact words. 100% of its quotes were found verbatim in the transcript, and its grades agree with the reference 83% of the time, against 72% for GLM 5.
vLLM keeps the transcript cached, so only the first question pays to read the call. Each one after takes a few hundredths of a second, even on a 90-minute call.
Drag to your own call volume. The defaults are the month the model card prices: 10,000 calls, with 10 questions and 10 checklists each.
100,000 decisions and 100,000 insights a month.
You save, against a frontier API
$101,820 a year
$8,485 every month · 42× cheaper
What each costs a month
Dwight runs on one H100 at about $3.95 an hour, billed per second, with the GPU off when idle. GLM 5 and the frontier model are at list price, and each decision question sends a transcript of about 10,000 tokens. Sending every question in one request cuts their decision cost about tenfold, and gives up one-token answers with a confidence.
Dwight is tuned for precision over coverage. When it flags something, it is more often right and always quoted exactly. The trade-off: it surfaces about half of what a frontier model finds, against about 61% for GLM 5. For full coverage, Zime runs larger models.
Dwight 4BJev 1.13
313 questions on 120 held-out calls. An independent LLM referee set the labels. Same questions for both models.
Dwight 4BGLM 5
475 call checklists on 40 held-out calls from 4 companies. The reference is a large frontier model on the same calls.
At scale, across tens of thousands of questions, Dwight agrees with a large frontier model 0.80 on next meeting, 0.76 on competitor, 0.75 on objections and 0.76 on pricing, with calibration error of 0.02 to 0.05.
Dwight is the small model we chose to give away. In production, Zime runs larger models built on the same sales data, and turns what they find into moves your reps actually make.
Free Apache 2.0, your GPU
For teams building their own call analytics who need a fast, cheap, honest answer on every call.
Every signal run and maintained by Zime
For revenue teams that need every signal in every call, and the move their best rep would make next.
Dwight is a focused 4B model, and open weights mean you run it. Here is what that puts on your side, and what to do about each.
It is conservative and finds fewer insights than larger models, most of all objections.
What to doDo not treat a missing insight as "it didn't happen". Use larger models where you need every signal.
It was trained on synthetic B2B software sales calls, and is weaker on companies and call types unlike them.
What to doTest it on a sample of your own calls before you rely on it.
English only.
What to doRoute calls in other languages somewhere else.
It reads 32k tokens, about a 90-minute call.
What to doShorten longer transcripts before you send them.
It is not a source of truth about people.
What to doCheck important outputs against the transcript. Every excerpt carries a timestamp to make that quick.
You host it: one 24 to 80 GB GPU, best with vLLM.
What to doTurn on prefix caching to get the follow-up speeds above, and plan for scaling and uptime.
Fine-tuned from Qwen3-4B-Instruct-2507 with LoRA on 8 H100s, on about 35,000 insight examples covering 500+ call behaviors and 8,000 decision questions, all from synthetic B2B software sales calls. The training data is not released.
Ask any question a person could answer from the call. Specific questions that say what each option means work best.
One question at a time. Read the option-token probabilities off the first generated token and renormalize over the options. That is the confidence.
system: You are a sales decision model. Read the conversation and answer the question with exactly one option token.
user:
<conversation>
[00:00:05] [Maya Brooks]: Can we meet Tuesday at 10?
[00:00:09] [Dan Ruiz]: Tuesday 10 works, I'll send the invite.
</conversation>
<question id="next_meeting_set" type="choice">
Was a concrete next meeting agreed, with a date and time?
</question>
Options:
A) High
B) Medium
C) Low
Answer with exactly one letter.Runs best on vLLM with enable_prefix_caching=True. Full runner: examples/decide.py
Dwight 4B is Zime's open-weights language model for sales calls. It has 4 billion parameters and reads a full call transcript, up to 32,768 tokens or about a 90-minute call. It does two jobs: it answers a typed question about the call (yes or no, a choice, or a 0 to 10 score) with one token and a calibrated confidence, and it writes structured insights as JSON with a timestamp and a verbatim quote.
About $205 a month for 10,000 calls with 10 decision questions and 10 insight checklists each, self-hosted on one NVIDIA H100 billed per second at about $3.95 an hour. The same month costs about $2,970 on GLM 5 and about $8,690 on a frontier API model at list prices, so Dwight is about 42 times cheaper than the frontier model. A dedicated always-on H100 costs about $2,900 a month and only pays off at much higher volume.
On one H100 with vLLM, the first question on a 20-minute call takes 0.13 seconds and each further question on the same call takes 0.02 seconds, because vLLM keeps the transcript cached. Ten questions finish in about 0.31 seconds. GLM 5 takes 1.9 seconds per question through its API. On a 90-minute call the first question takes 1.06 seconds and later questions stay under 0.08 seconds.
On 313 questions across 120 held-out calls, scored by an independent LLM referee, Dwight scored 0.79 on next meeting set against 0.57 for Jev 1.13, and 0.86 on pricing sentiment against 0.74. Jev led on unresolved objections, 0.79 to 0.64. Across tens of thousands of questions, Dwight agreed with a large frontier model 0.75 to 0.80 of the time, with calibration error (ECE) of 0.02 to 0.05.
In the published comparison, 100% of Dwight's quotes were found word for word in the transcript, against 98.7% for GLM 5, and its grades agreed with the reference 83% of the time against 72%. Dwight is tuned for precision over coverage: it surfaces about half of the insights a frontier model finds, against about 61% for GLM 5.
One GPU with 24 to 80 GB of memory. The weights are about 8 GB in bf16. It runs best on vLLM with prefix caching turned on. On one H100, running four checklists at a time, it writes about 2,000 insight checklists an hour.
The weights are released under the Apache 2.0 license on Hugging Face at huggingface.co/zime-ai/dwight-4b. Dwight is fine-tuned with LoRA from Qwen3-4B-Instruct-2507, on about 35,000 sales-call insight examples covering 500+ call behaviors and about 8,000 decision questions, from synthetic B2B software sales calls. The training data is not released.
It is conservative and finds fewer insights than larger models, most of all objections. It is weaker on companies and call types unlike its training data, is English only, and reads at most 32k tokens, so longer transcripts must be shortened. It is not a source of truth about people: check important outputs against the transcript.
Dwight is the small model Zime chose to release openly, built from what Zime learned analyzing sales conversations. In production, Zime runs larger models that find more of what a call holds, and turns them into Behaviors through the Behavior Graph, delivered in Slack, Microsoft Teams and Claude.
Run Dwight on your own GPU for a fraction of a cent a question. When you want the full picture of every deal, Zime runs the bigger models for you.