Kronos vs Qwen3-4B-Instruct-2507

Kronos

Forecasts the next price bars from a history of up to 512 candlestick bars. Every size fits in under 0.5 GB and runs on a CPU, but only through the publisher's own Python package. Nobody has benchmarked it independently, its quickstart data was deleted from the repository, and the publisher hasn't replied since November 2025.

Qwen3-4B-Instruct-2507

A small, fast cousin for instruction following, short tool-calling jobs and long documents on modest hardware, with no thinking step. It runs in about 3 GB at 4 bits on every common runtime, and two providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

Kronos

1. Clone https://github.com/shiyu-coder/Kronos and pip install -r requirements.txt, then load Kronos.from_pretrained("NeoQuasar/Kronos-small") and KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base") and wrap them in KronosPredictor
2. Pair Kronos-mini with Kronos-Tokenizer-2k and Kronos-small or Kronos-base with Kronos-Tokenizer-base, the cards' pairing
3. The examples read examples/data/XSHG_5min_600977.csv, which no longer exists; point them at tests/data/regression_input.csv, 2,500 five-minute bars
4. Keep the lookback at or under max_context, 512 bars, and pass the future timestamps you want as y_timestamp, since the calendar fields come from them
5. Raise sample_count for a steadier forecast, but expect only the mean back; call the lower-level auto_regressive_inference yourself if you need the paths
6. Treat forecasts as model output to check, not signals; the README says its own backtest pipeline isn't a production trading system

Qwen3-4B-Instruct-2507

1. Quickest start is ollama run qwen3:4b-instruct, 2.5 GB. Plain qwen3:4b is the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes
3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
6. Drop the context to 32,768 tokens if memory runs out, as the card suggests
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented