Qwen-Code v0.10.5 夜間版本發佈
💡Qwen-Code nightly v0.10.5 bump—test latest for code gen stability gains.
⚡ 30-Second TL;DR
有什麼變化
夜間構建 v0.10.5-nightly.20260222.fc04ba1e
為什麼重要
此次輕微夜間更新維持 Qwen-Code 開發動能。為使用者提供 0.10.5 版本穩定性修復的早期存取。廣泛影響低，但對積極測試者有用。
下一步行動
Clone the qwen-code repo and test v0.10.5-nightly build for coding task improvements.
關鍵要點
- •夜間構建 v0.10.5-nightly.20260222.fc04ba1e
- •透過 @qwen-code-ci-bot 將版本提升至 0.10.5
- •變更提交於 PR #1886
- •提供自 v0.10.5 的完整變更日誌
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 6 個來源。
🔑 增強重點摘要
- •Qwen-Code v0.10.5-nightly is a terminal-based open-source AI coding agent optimized for Qwen models, with the latest nightly build released on 2026-02-22 featuring a version bump in PR #1886.
- •Qwen-Code supports advanced coding workflows including Slash Commands, MCP, PlanMode, TodoWrite, SubAgent, Multi Model support, and tools like WebFetch, Bash, and FileRead/Write.
- •This release aligns with Alibaba's Qwen team's recent launch of Qwen3-Coder-Next, an 80B MoE model with 3B active parameters designed for coding agents and local development.
- •Qwen3-Coder-Next, released in February 2026, achieves high coding performance on consumer hardware (e.g., 64GB MacBook or RTX 5090) with 256K context length and Apache 2.0 license.
- •Qwen-Code is part of the broader Qwen3.5 series on GitHub, focused on understanding large codebases and automating development tasks.
📊 競品分析▸ Show
| Feature | Qwen-Code v0.10.5 | Qwen3-Coder-Next | Claude Sonnet 4.5 (Benchmark Reference) |
|---|---|---|---|
| Parameters | N/A (Agent Tool) | 80B total, 3B active (MoE) | N/A (Closed) |
| Context Length | N/A | 256K tokens | High (Proprietary) |
| Hardware Req. | Terminal-based | 42GB+ RAM (consumer GPUs) | Cloud-only |
| License | Apache 2.0 (inferred) | Apache 2.0 | Closed |
| Benchmarks | N/A | Sonnet 4.5-level coding | Baseline for comparison |
| Pricing | Free/Open-source | Free/Open-weight | Paid API |
🛠️ 技術深入
- Qwen-Code is a terminal AI agent with roadmap features: Coding Workflow (Slash Commands, MCP, PlanMode, TodoWrite, SubAgent, Multi Model, Chat Management), Tools (WebFetch, Bash, TextSearch, FileRead/Write).
- Related Qwen3-Coder-Next architecture: Hybrid Gated DeltaNet + MoE + Gated Attention; 512 experts, 10 activated per token; trained via large-scale executable task synthesis + RL.
- Qwen3-Coder-Next specs: 80B total parameters, 3B activated, 256K context, causal LM, supports tool calling (JSON), runs at 20-40 tokens/sec on quantized consumer hardware (e.g., RTX 5090, AMD Radeon 7900 XTX).
- Available in GGUF and MLX formats for local inference via LM Studio or agentic harnesses like Codex, Claude Code, OpenCode.
- Qwen3.5 GitHub repo: Apache 2.0, optimized for Qwen models in codebases.
🔮 前景展望AI analysis grounded in cited sources
Qwen-Code v0.10.5 nightly advances open-source terminal-based coding agents, enhancing local AI development with Qwen3-Coder-Next's efficient MoE design, potentially democratizing high-performance coding on consumer hardware and competing with cloud-dependent models.
⏳ 時間線
📎 來源 (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： Qwen (GitHub Releases: qwen-code) ↗
每週 AI 簡報
每週一封，可隨時退訂。