LANGUAGE MODEL Alibaba Last updated:

Qwen 2.5

Open Frontier-Comparable from Alibaba

Alibaba's Qwen 2.5 generation, released September 2024, was the first Chinese open-weight model family to exceed Llama 3.1 on most benchmarks at comparable size. The release covered seven sizes (0.5B → 72B) plus specialized variants for code (Coder) and math, all under Apache 2.0. By late 2024, Qwen 2.5 had overtaken Llama as the most-downloaded model family on Hugging Face globally.

Intelligence
Good
Speed
Slow
47 tok/s output
Cost
Low
$0.36 in / $0.40 out
Context
131K
Up to 131,072 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/M out).
Intelligence
  • Top 1%≤ 1%
  • Top 5%≤ 5%
  • Top 10%≤ 10%
  • Good≤ 25%
  • Medium≤ 50%
  • Below avg> 50%
Speed
  • Top 1%≥ 345 tok/s
  • Top 5%≥ 237 tok/s
  • Top 10%≥ 196 tok/s
  • Good≥ 146 tok/s
  • Medium≥ 90 tok/s
  • Slow< 90 tok/s
Cost
  • Freeopen weights · self-host
  • Low< $1 / M out
  • Moderate$1–5 / M out
  • High≥ $5 / M out

Why it matters

Qwen 2.5 made it impossible for any serious enterprise procurement decision in 2025 to skip the open-Chinese-model option. Even US enterprises with concerns about Chinese-origin models routinely benchmark against Qwen 2.5 / 3 to negotiate better pricing from OpenAI / Anthropic. The market structural effect is large.

Core Capabilities

Long Documents
Handles entire codebases, books, and multi-doc RAG.
Generative
Produces images, video, audio, or other media.
Agent Workflows
Built for tool use and autonomous tasks.

Context Window

131k tokens
≈ 101 pages
4k Chat 聊天
32k Long docs 长文档
128k This model 本模型
400k Multi-doc 多文档
1M Codebase 整个代码库
10M

Availability

API
Not available
Product / App
Not available
Open Source
Released
Enterprise

Pricing Model

Free / self-host
Open weights — pay only for compute.
Self-host

Capability / Performance

Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).

Lower 20% Upper 80% This model
Reasoning
AA Intelligence Index · scaled to 10
1.7
5.6
4.9
Coding
SciCode · scaled to 10
1.8
4.3
4.3
Agentic tasks
Terminal-Bench Hard · scaled to 10
0.2
3.6
2.6
Context / memory
Context window size · log-scaled
6.0
9.0
6.1
Cost efficiency
Input price ($/M tokens) · cheaper scores higher
6.2
10.0
10.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.

What it feels like

  • First open-source model to top OpenCompass leaderboard — beat Llama 3.1 405B with 1/5 the parameters
  • 72B-Instruct: 74.2 on coding, 77 on math — outscored Claude 3.5 Sonnet (72.1) and GPT-4o (70.6) at release
  • MMLU 85+, HumanEval 85+, MATH 80+ — frontier-tier across the board for an open-weights model
  • 100+ models across the family (sizes from 0.5B to 72B, plus VL/Audio/Coder variants) — full ecosystem
  • Pre-training jumped from 7T to 18T tokens with heavy emphasis on knowledge / coding / math
  • Apache-2.0 license + competitive quality made it the de-facto open-source default in late 2024

Best use cases

  • Self-hosted production where Llama 3.1 405B is too big to serve
  • Multilingual deployments (29 languages with strong coverage)
  • Fine-tuning on private code or technical corpora
  • Distillation into smaller production models (the 7B/14B variants are excellent students)

Tools to try

Not ideal for

  • Frontier reasoning by mid-2025 — Qwen 3 / DeepSeek R1 / Claude 4 series have moved past it
  • Vision tasks (use Qwen 2.5-VL) or audio (use Qwen 2.5-Audio)
  • Workflows where Western-platform compliance is contractual

Model Evolution

qwen is Alibaba's language model family.

View full evolution tree →