🕷️
SetupAI
🧧Stalecollected in 15h

Qwen-Code v0.12.3 Nightly Released

Qwen-Code v0.12.3 Nightly Released
PostLinkedIn
🧧Read original on Qwen (GitHub Releases: qwen-code)
#nightly-build#github-release#changelogqwen-code

💡Latest Qwen-code nightly fixes could boost your coding AI performance—check changelog now

⚡ 30-Second TL;DR

What Changed

Nightly release: v0.12.3-nightly.20260314.f1ee4638

Why It Matters

This minor nightly update likely includes bug fixes and optimizations for Qwen-code, benefiting developers using it for coding tasks. It maintains momentum in open-source code LLMs.

What To Do Next

Review the changelog on Qwen-code GitHub and test the new nightly build in your coding workflows.

Who should care:Developers & AI Engineers

🧠 Deep Insight

Web-grounded analysis with 6 cited sources.

🔑 Enhanced Key Takeaways

  • •Qwen-Code is a developer tool with a roadmap featuring experimental LSP support in v0.7.0, Anthropic API provider integration, and concurrent runner for batch CLI execution with Git in v0.6.0[3].
  • •Qwen3-Coder-Next, released by Alibaba's Qwen team in February 2026, uses a Mixture-of-Experts (MoE) architecture with 80B total parameters but only 3B activated per inference, enabling local runs on consumer hardware like RTX 5090[2][4][5].
  • •Qwen coding models excel in benchmarks, with Qwen3 achieving 69.6% on SWE-Bench Verified and 88.4% on HumanEval (7B version), outperforming GPT-4 in some metrics, plus support for 92 programming languages[1].
📊 Competitor Analysis▸ Show
Feature/BenchmarkQwen3-Coder-NextGPT-4Claude Opus 4DeepSeek V3.1
SWE-Bench Verified/Pro69.6% / 44.3%[1][2]87.1% (HumanEval)[1]Outperformed on Tau2-Bench[1]Outperformed on Tau2-Bench[1]
Active Parameters3B (80B total MoE)[2]N/A (proprietary)N/A (proprietary)N/A
Context Length256K[1][2]N/AN/AN/A
PricingFree (open weights, local)[1][2]API costsAPI costsN/A

🛠️ Technical Deep Dive

  • •Architecture: Hybrid Gated DeltaNet + MoE + Gated Attention, with 512 experts total and 10 activated per token[2].
  • •Training: Large-scale executable task synthesis combined with reinforcement learning (RL)[2].
  • •Context: Native 256K tokens, efficient VRAM usage avoiding ballooning in traditional SDPA models[1][4].
  • •Additional features: Hybrid reasoning (thinking/non-thinking modes), MCP for agent integration, trained on 36 trillion tokens[1].

🔮 Future ImplicationsAI analysis grounded in cited sources

Qwen-Code nightly releases will accelerate integration of advanced MoE models into local coding tools
The v0.12.3 nightly aligns with Qwen3-Coder-Next's February 2026 release, incorporating efficient 3B active parameter MoE for consumer hardware[2][5].
Open-source Qwen models will capture larger share of coding agent market by 2026 end
Benchmarks show Qwen3 leading free coding AI with top scores on SWE-Bench and HumanEval, plus zero API costs and privacy advantages over proprietary rivals[1][6].

⏳ Timeline

2026-02
Qwen3-Coder-Next released as open-weight MoE model for coding agents[2][5]
2026-03
Qwen-Code v0.12.3-nightly.20260314.f1ee4638 released with changelog from v0.12.3[]
2025-12
Qwen-Code v0.7.0 adds experimental LSP service and Anthropic provider support[3]
2025-11
Qwen-Code v0.6.0 introduces concurrent runner with Git integration[3]
2025-10
Qwen-Code v0.2.0 integrates ACP/Zed editors[3]
2025-09
Qwen-Code v0.1.0 launches with terminal UI, OpenAI protocol, MCP, and multi-model support[3]
🧧Read original article on Qwen (GitHub Releases: qwen-code)
📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #nightly-build

Same product

More on qwen-code

Same source

Latest from Qwen (GitHub Releases: qwen-code)

AI-curated news aggregator. All content rights belong to original publishers.
Original source: Qwen (GitHub Releases: qwen-code) ↗