Qwen3 4B Instruct 2507

4B parameter open-weight model. This entry is Qwen3-4B-Instruct-2507, the non-thinking update of Qwen3 4B. It is a different checkpoint from the Qwen3 4B entry, which is Qwen3-4B-Thinking-2507 (Ollama's qwen3:4b tag). Facts below come from the publisher's model card and config.json (Qwen/Qwen3-4B-Instruct-2507, read 2026-10-27): 4.0B parameters (3.6B excluding embeddings), 36 layers, 32 attention heads with 8 key-value heads, a 262,144-token native context, Apache 2.0. The card states it supports only non-thinking mode and never writes a think block, and lists gains over the original Qwen3-4B non-thinking mode in instruction following, logical reasoning, text comprehension, mathematics, science, coding, tool use, multilingual long-tail knowledge and 256K long-context understanding. Qwen recommends Temperature 0.7, TopP 0.8, TopK 20 and MinP 0, an output length of 16,384 tokens for most queries, and a shorter context such as 32,768 if memory runs out. The memory figures are the file sizes of its three explicit Ollama tags (registry manifests checked 2026-10-27), not measurements, and none of the three has been run on our hardware. Weights are only part of the memory bill: from the config, the KV cache at FP16 is about 0.15 MB per token, so roughly 0.6 GB at a 4,096-token context and 4.8 GB at 32,768 (computed, not measured). The Q4 weights are 2.5 GB, so a 4 GB card runs it at a short context only.