@mylifcc: They took 2.3 million pieces of 'thinking process' data left by Claude (a very powerful and expensive AI) during inference, and used it to train a very small model (Qwen3-4B, only 4 billion parameters). As a result, this small model performed very 'consistent' in tests: In 512 different tests, the output was exactly the same every time...
Summary
Using 2.3 million pieces of Claude inference trajectory data to distill the Qwen3-4B model, resulting in a small model with 100% output consistency and extremely low hallucination. Moreover, the student model was not limited by the teacher model, and it also converged to a universal truth.
View Cached Full Text
Cached at: 07/04/26, 06:39 AM
They took 2.3 million “thinking process” traces left by Claude (a very strong and expensive AI) during inference and used them to train a very small model (Qwen3-4B, only 4 billion parameters).
As a result, this small model performed remarkably “stable” in tests:
- In 512 different tests, its output was exactly the same each time (100% consistent).
- The output had almost no randomness, essentially deterministic.
- Almost no nonsense (very low hallucination).
- Surprisingly, the small model was not limited by the large model (the teacher), but instead excelled at reliability.
- Most absurdly, it even summarized a “universal truth” — the “Egypt wins” that everyone was laughing about in the comments.
ali (@waterloo_intern): we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B
- 100% self-consistency @ 512 samples
- 0.00 bits output entropy
- zero hallucination variance
turns out the student is not bounded by the teacher. it also converged on one universal truth.
we open-sourced the
Similar Articles
@WolfTrainer_101: A 9B super reasoning model suitable for local deployment by security researchers - Qwythos-9B-Claude-Mythos-5-1M. Based on the Qwen3.5-9B base model, it is further trained with 500 million+ high-quality Claude Mythos reasoning traces. Core highlights: 1. Native 1M…
Qwythos-9B is a 9B super reasoning model based on the Qwen3.5-9B base, further trained with 500 million+ Claude Mythos reasoning traces. It natively supports 1M long context and tool calling, designed for local deployment by security researchers. It significantly outperforms the original base model on MMLU and mathematical reasoning.
@zhixianio: After receiving the new machine, I began an 'ascetic' practice of forcing myself to use local models for common tasks. I thought it would be painful, but both speed and quality greatly exceeded my expectations: Model: Qwen3.6-35B-A3B-oQ6-fp16-mtp, Running: oMLX, with N…
The author uses the Qwen3.6-35B-A3B model and oMLX tool on the new local machine for daily tasks, finding that both speed and quality far exceed expectations, even outperforming remote LLMs in PA and coding scenarios, demonstrating a significant improvement in on-device AI capabilities.
@RookieRicardoR: Domestic models break through again, matching top models like Claude 4.6 and Gemini 3.1 Pro. Just tested Qwen3.7-Max, sharing some real thoughts. Last night I topped up as soon as the API went live and chose three tasks (see video) to test Qwen3.7-Max's frontend capabilities…
The user tested Qwen3.7-Max and believes it matches top models like Claude 4.6 and Gemini 3.1 Pro in frontend, computing power, and Agent capabilities. Its reasoning ability has significantly improved, and with monthly iteration speed, it has become a first-tier domestic model.
@sanbuphy: K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac, using the niche Zig language to implement and optimize inference, demonstrating the new model’s generalization ability. After 4,000+ tool calls and 12+ hours of continuous operation, K2.6 iterated 14 times…
K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac, using the niche Zig language to implement and optimize inference, demonstrating the new model’s generalization ability. After 4,000+ tool calls and 12+ hours of continuous operation, K2.6 iterated 14 times, boosting throughput from ~15 tokens/s to ~193 tokens/s, ultimately achieving 20% faster inference than LM Studio.
@ZeroZ_JQ: https://x.com/ZeroZ_JQ/status/2079504278922891568
This article provides an in-depth explanation of the principle of reasoning effort in large language models: by adjusting the reasoning effort level (low/medium/high) on the same model, more intermediate reasoning trajectories (serial autoregressive computation) are allowed before outputting the final answer, thereby making trade-offs among answer quality, response speed, and computational cost. The key is that the model has learned to organize longer generation processes into effective computation, rather than simply increasing parameters or network layers.