Qwen3-4B

QLoRA introduces 4-bit quantization technology, making it possible to fine-tune models with 65 billion parameters on a single RTX 3090/4090 GPU with 24 GB of VRAM.

The off-policy distillation experiment uses Qwen3 0.6B as the student model and Qwen3 4B as the teacher model.

VASPO improves average accuracy by 2.59 percentage points over PPO on Qwen3-4B and by 1.81 percentage points on OLMo3-7B.

reef's OpenClaw-RL recipe uses 72 GSM8K problems as the task stream, with Qwen3-4B serving simultaneously as the policy model and the process reward model, Qwen3-32B as the student model, using seven GPUs, and setting a threshold that must be reached before an update is accepted.

On Qwen3-4B, after pruning MLP layers to 2:4 sparsity and adding rank-64 correction, CACTUS combined with LoGRA still maintains 0.711 cosine alignment with the C4 gradients of dense LoGRA after projection.

When doing 50% pruning on Qwen3-4B, CACTUS raised the compressed-dense per-sample gradient cosine similarity on WikiText-2 from 0.670 to 0.755.

When doing 50% pruning on Qwen3-4B, CACTUS reduced the gap between C4 perplexity and the dense model by 45%.

Gemma 4 E2B's KV cache capacity is 321,280 tokens, which is 2.6 to 6.3 times that of Granite 3.3 2B (125,184), SmolLM3-3B (120,320), Qwen3-1.7B (100,096), and Qwen3-4B (50,688).

On the GSM8K mathematical reasoning test, Qwen3-4B scores 0.915 and Gemma 4 E2B scores 0.911, a difference of +0.4 percentage points with a confidence interval crossing zero, making it the only model on par with Gemma.

On the BFCL tool-calling test, SmolLM3-3B scores 0.945, slightly above Gemma 4 E2B's 0.927, but the difference is not significant; Qwen3-4B (0.940), Qwen3-1.7B (0.935), Qwen2.5-1.5B (0.910), and Granite 3.3 2B (0.920) are all at the same level as Gemma.

In the scenario where 16 requests each carry a 3,072-token long prompt, Gemma 4 E2B maintains 1049 tokens/second, while Granite 3.3 2B drops to 732 (0.70x), SmolLM3-3B drops to 725 (0.69x), Phi-4-mini drops to 404 (0.39x), and Qwen3-4B reaches only 373 (0.36x).

When using SGUID, Qwen3-4B improved by 0.5 percentage points on HMMT25 after the first round, and by a cumulative 1.1 percentage points after the second round.

In the Qwen3-4B comparative experiment, updating the model directly with unfiltered skills without using SGUID filtering decreased performance by 0.3 percentage points on HMMT25.

The paper evaluates Qwen3.6-27B on five competition tasks from MLE-Bench Lite and Qwen3-4B on the text adventure game Zork I (Jericho framework).

Using Qwen3-4B as the base model, applying supervised fine-tuning with only counterexample data causes the model's ability to recognize true theorems to collapse from 0.27 to 0.00.

The collapse in true-theorem recognition ability caused by SFT was observed across four repeated experiments with different random seeds and was reproduced on Gemma-3-4B.

Using a 4B model to rerank the candidate shortlist filtered by BM25 outperformed BM25 alone (+0.047, confidence interval [0.020, 0.073]), even though the 4B reranker itself did not pass qualification.

LessThink-Qwen3-4B is a model obtained by post-training Qwen3-4B to reduce token consumption by approximately 44% during the inference stage while preserving the original knowledge coverage and response style as much as possible.

On the Qwen3-4B model, 76.9% (±2.8%) of stated reasoning steps at the most sensitive middle layers of the network were causally load-bearing.

The random-position null baseline for Qwen3-4B was only 11.3%, and the upper bound for directly patching lower-level prompt facts was 83%.