pulse-local-qwen3-0.6b

Qwen3-0.6B-pulse

A 0.6B Qwen3 tuned for one narrow job inside Pulse Local, a budgeting app for the Mac where your transactions, the models and every question stay on your machine. The app downloads this file the first time it starts.

This is not a general assistant, and on its own it is not a good one. In the app it does a small part of the work. Code answers every question about your records that it understands, and computes every figure. This model words the questions code cannot answer, such as explanations and requests for a judgement, and it is tuned above all to describe your situation without telling you what to buy or sell.

What it does in the app

The app answers a question in this order.

1. Code reads the question into a calculation: what to count, about what, over which dates. If it understands the question, it runs the calculation against your records and writes the answer from a fixed set of sentences. Most questions end here, and the model is never asked.
2. A second model checks every figure. Since October 2026 the app also reads each question with Qwen3-4B-Instruct-2507 (unsloth's Q4_K_M, about 2.5 GB, on by default and switchable off in Settings). A figure is shown only when the two readings lead to the same calculation. Otherwise the app says it is not sure. This model plays no part in that check.
3. Written notes answer questions about the app itself and common finance terms.
4. This model answers what is left. Its answers are labelled in the app, every money figure in them must come from the question or from a lookup made for that answer, and an answer that names a financial product the question did not name is replaced with a written one.

The embedding model used for routing is bge-small-en-v1.5 (GGUF from CompendiumLabs), downloaded alongside this one.

How the app is tested

We test the app, not this model alone, because the app is what people use. Each test uses three invented households written by a new AI author that has never seen the app's code or any earlier test, about 55 questions per household, roughly half of them follow-up conversations. A separate program computes every expected answer from the households' transactions, every figure in every reply is checked against it, and a person reads every answer.

With the check, the app answers a little under half the questions correctly and says it is not sure to most of the rest. These figures describe the whole app, not this model.

What this model is good and bad at

Measured with the app's tools but no code in front of it, on 77 everyday budgeting questions, counting an answer with no figure as a miss:

The 8B often described the lookup it would make instead of making it, and asked the user for figures the records already held. This model makes the call. Neither is reliable enough to answer money questions unsupervised, which is why the app does not ask them to.

What the tuning changed

Posture is the rule the app exists to keep. Pulse describes a financial situation and never tells anyone to buy, sell, hold, trim or allocate. The base model broke it, answering "selling now could be a good move" to a question about a falling holding. This tuning exists to fix that.

Files

Q6_K was measured and was no better than Q4_K_M in the app, so it is not published.

Running it

• Chat template. Use Qwen3's, without a prefilled empty thinking block. The model was trained on a template that emits no <think> block, and a prefilled one makes it drift.
• Tool calls. Use a plain generation loop that parses <tool_call>{"name": ..., "arguments": {...}}</tool_call> from the text. The built-in function calling in node-llama-cpp drops this model's calls, while the same GGUF works in llama.cpp and MLX. The app uses its own loop for this reason.
• Sampling. Use greedy decoding, temperature 0. Sampled, it can break off in the middle of a tool call.
• Context. 4096 tokens is enough for everything the app sends it.

Training

• Base: Qwen/Qwen3-0.6B, Apache 2.0.
• Method: LoRA, rank 16 on 16 layers at learning rate 3e-5, with mlx-lm on an M4 Mac, then fused and converted to GGUF with llama.cpp.
• Data: about 1,860 synthetic rows, roughly 1,700 written answers plus tool-call rows generated from the app's own tool schemas and calculator. The answers were distilled from gpt-4.1-mini and filtered for directive phrasing, boilerplate, em dashes and length. No personal or user data was used at any point. Every figure in the training data is synthetic.
• A later round on 5,617 rows was trained and measured. It came out flat, 8 items better and 9 worse out of 348, so it is not published.

Limitations

• English only, and US personal finance only.
• No internet and no live data. It is trained to say so rather than invent a price, a rate or an index level.
• It is poor at one-word yes or no answers, 13 right out of 30 on questions with a clear answer, with a lean toward yes. The app never asks it for one.
• It gets the method of a calculation right and the numbers wrong, which is why the app computes every figure in code.
• It is not a financial adviser, and neither is the app.

Where it runs

The app that carries this model is at pulseos.finance, for Apple Silicon Macs, signed and notarised. It is pay what you want, including nothing, and the weights here are open under the license below. The model itself has not changed since September 2026. The app around it has changed a great deal, which is why the app's results keep moving while this model's do not.

Provenance

The training answers were generated with OpenAI models. If you build on this model, check OpenAI's terms for your own use.

License

Apache 2.0, inherited from the base model, whose LICENSE is included.