Qwen2.5-14B-Instruct
#56 v Open-source jazykové modelyalibaba · v2.5 · 14b instruct · od 2024-09-19 · 2× · naposledy 30. 6. 2026
7
Momentum
Qwen2.5-14B-Instruct is an instruction-tuned, dense language model from Alibaba Cloud's Qwen team with 14.7 billion parameters, released in September 2024 under the Apache 2.0 license. It natively supports a context length of up to 128,000 tokens and can generate up to 8,000 tokens. The model was pre-trained on 18 trillion tokens and covers over 29 languages. According to the official technical report, it performs comparably to GPT-4o-mini across several benchmarks.
Vývoj momenta
04.04.03.07.
Vlastnosti
| Benchmark Score (MMLU/Similar) | MMLU: 79.7; BBH: 78.2 (Qwen official blog); MMLU-Redux: 80.0%; GSM8k: 94.8%; MATH: 80.0%; HumanEval: 83.5% (llm-stats.com) |
| Context Window | 128,000 tokens (native support); default config.json set to 32,768 tokens; output up to 8,000 tokens |
| Model Size (Parameters) | 14.7 billion parameters (active and total); trained on 18 trillion tokens |
| Memory Requirement | BF16 (full precision): approx. 29.6 GB VRAM (model weights); Q4_K_M quantization: approx. 8.7 GB; Q8_0: approx. 14.7 GB (each plus 1–2 GB KV cache overhead) |