Qwen3.5-Coder

• 1. overview (603 chars)

  Qwen3.5-Coder is a code-specific open weight model series launched by the Alibaba Tongyi Qianwen team based on the Qwen 3.5 family. It is positioned in the same vein as Qwen3-Coder and Qwen3-Coder-Next, but is bound to the Qwen 3.5 hybrid architecture released in February 2026. Two sizes have been disclosed so far: 7B and 32B, both natively support 256K context, open weight, Apache 2.0 license, commercially available, and are oriented to local deployment and software engineering agent (Coding Agent) scenarios. Its core selling point is very straightforward - using open source weight to stuff "warehouse-level programming + tool calling + long context" into a volume that can run on consumer-grade hardware.

• 2. Product background (909 chars)

  Qwen is a large open source model family created by Alibaba Cloud since 2023. By 2026, it has become one of the most downloaded open source model series in the world. On February 16, 2026, Alibaba officially released Qwen 3.5 (flagship Qwen3.5-397B-A17B), which adopts Gated DeltaNet linear attention and sparse MoE hybrid architecture, supports 201 languages, native multi-modality, and up to 1M context. Qwen3.5-Coder is a specialized branch of this generation of architecture in the "code" dimension. According to LLM Reference, the release time of the two variants of Qwen3.5-Coder (7B, 32B) is marked as January 2026, earlier than the February release of the Qwen 3.5 flagship. The timeline is slightly misaligned, but both belong to the 3.5-generation technology stack. It is worth mentioning that the official Qwen 3.5 main blog does not list the "Coder" sub-model separately, but emphasizes Qwen Code, a command-line programming tool; the independent Qwen3.5-Coder information mainly comes from the coder page of the Qwen official blog and the Hugging Face model set.

• 3. Product functions and usage experience (893 chars)

  The specifications of the two open weight variants are clear: Qwen3.5-Coder-7B (2026-01, 256K context, 7B parameters) and Qwen3.5-Coder-32B (2026-01, 256K context, 32B parameters), both marked as supporting code execution. It continues the two traditional advantages of the Qwen code model: first, the native 256K long context, which can read the entire code warehouse at once, and is suitable for warehouse-level refactoring and cross-file dependency analysis; second, it is Agent-friendly and can cooperate with Qwen Code, Claude Code, Cline, Kilo, Trae and other tools to create an autonomous closed loop of "writing code - running tests - reading errors - fixing bugs - verifying". Coder variants (such as mdq100/qwen3.5-coder:35b) based on the official Ollama version of Qwen3.5 have also appeared in the community. By lowering the temperature and adding a presence penalty, the output is more certain, and the visual input and tool calling capabilities are retained, making it easy to use out of the box in local programming tools such as OpenCode.

• 4. Pricing and business model (716 chars)

  As an open weight model, Qwen3.5-Coder itself is free, Apache 2.0 is commercially available, and developers can directly download the weights for local deployment in Hugging Face / ModelScope. If you use the managed API, you can reuse Alibaba Cloud Bailian's Qwen 3.5 series billing: Qwen3.5-Plus (1M context, built-in tools) is priced on a tiered basis in mainland China, about 0.8 yuan/million input tokens (0–128K), 4.8 yuan/million output tokens (0–128K), and there is an additional price for long contexts and thinking modes; there is also the Alibaba Cloud hosted version Qwen3.5-Plus. There are multiple acquisition paths with community routing (Together, OpenRouter, Ollama). Compared with closed-source flagships, the greatest value of open-source weight is "one download, permanent ownership, and data does not leave the domain."

• 5. User reviews (928 chars)

  Positive feedback focuses on three points: First, the cost structure - open source is free, can be run locally, and is extremely friendly to individuals and small teams with limited budgets; second, the native optimization of Chinese and Chinese-English mixed code scenarios, which has been repeatedly mentioned in the Chinese developer community; third, the warehouse-level understanding brought by long context makes multi-file reconstruction and PR analysis no longer a toy-level demonstration. Negative and reserved voices point to: Compared with Qwen 3.5 flagship and ultra-large MoEs such as Qwen3-Coder-480B, 7B/32B Coder variants still have a gap in complex multi-step Agentic coding; the benchmark of the community Coder parameter version (such as: 35b) comes from the basic Qwen3.5 model rather than independent evaluation, and there is a suspicion of "intergenerational borrowing scores"; some third-party pages compare the Coder variant released in January with Coder variants. The Qwen 3.5 flagship released in February is confused, which can easily mislead the selection.

• 6. Industry evaluation (641 chars)

  On SWE-bench Verified, the golden benchmark for programming agents, Qwen 3.5 flagship (397B-A17B) reached 76.4, approaching Claude 4.5 Opus (80.9) and GPT-5.2 (80.0), and significantly ahead of the previous generation Qwen3-Max-Thinking (75.3). The benchmarks published by the community Coder variant (mdq100/qwen3.5-coder:35b) are SWE-bench Verified 69.2, LiveCodeBench v6 74.6, and CodeForces score 2028, which is in the first echelon of open source code models. Horizontally, it faces a number of efficient open source competing products such as DeepSeek-V3.2, Kimi-K2, Qwen3-Coder-Next (only 3B activation reaches 70.6%), and its advantage lies in the coherence of "Qwen 3.5 new architecture + official ecosystem (Qwen Code, Bailian API)".

• 7. Controversies and Risks (942 chars)

  The first risk is the varying credibility of information sources. The authoritative disclosure of Qwen3.5-Coder is relatively thin: the official Qwen 3.5 main blog does not list this model separately, the independent coder page and the Hugging Face collection are the only primary sources, and the specific benchmark scores of 7B/32B mostly come from third-party aggregation sites and community reference versions, not Ali’s official unified evaluation. Secondly, the release time mark (2026-01) is earlier than the Qwen 3.5 flagship (2026-02), which means that "intergenerational belonging" and "whether the weight is fully open" still need to be based on the official model card. Finally, there are common concerns about open source code models - the hardware threshold for local deployment (32B quantification still requires tens of GB of video memory), the inference speed and memory usage under long contexts, and the real gap with closed-source APIs in ultra-complex long-range tasks. Developers need to conduct actual testing and verification when selecting models instead of just looking at the list.

• 8. Applicable groups and usage suggestions (862 chars)

  Who is it suitable for: Teams who want to keep their code out of the domain and need local or private deployment; Chinese/Chinese-English mixed development scenarios; budget-sensitive individual developers and small companies who want to use Apache 2.0 commercially but are unwilling to pay for closed-source APIs; engineering teams that do Coding Agents, IDE plug-ins, and DevOps automation. Who it is not suitable for: Users who pursue extremely complex multi-step Agentic coding and can accept the cost of cloud (should go directly to Qwen 3.5 flagship or Claude/GPT closed source flagship); enterprises that require official SLA and commercial support and have low trust in the "community debug version". Alternatives: Qwen3-Coder-Next (3B activation, SWE-bench 70.6%, ultimate cost-effectiveness) and Qwen3-Coder-480B (API to ultra-large MoE) are available for the same generation; DeepSeek-V3.2 and Kimi-K2 are available for cross-family; Claude Sonnet / GPT-5 series are available for closed source.

• 9. Summarize (547 chars)

  Qwen3.5-Coder is an open source extension of the new architecture of Qwen 3.5 in the code dimension. It uses two open weight sizes of 7B/32B to package "long context + code execution + agent friendliness + Apache 2.0 commercial use" for local deployment scenarios. It is especially friendly to Chinese developers and privacy-sensitive teams. Its true level needs to wait for Alibaba’s official model card to be confirmed with the unified benchmark, but the direction is already clear: the open source code model is turning “warehouse-level programming” from a cloud luxury product into a compliant component that can be installed on your own machine.

No in-depth report yet

• 

  孔梅

  —

  这两天试了下 Qwen3.5-Coder 32B 量化版，跟 Cline 配合做代码审查，比之前用的 Qwen2.5-Coder 准了不少，尤其是跨文件重构的时候上下文保持能力提升明显。不过推理速度确实比预期的慢，同样硬件上比 Qwen3-Coder-Next 慢了将近一倍，期待后续优化版本能把推理效率提上来。

---

• 

  TheAnastasijaLazić_dev

  —

  256K 上下文确实不是虚标，我把整个项目丢进去让它分析依赖关系，效果比预期好。

---

• 

  6x7ghcy99

  —

  卡成 PPT……32B 量化版在 16GB 显存上跑推理速度不到 8t/s，还是得换 7B。

---

• 

  realKathyShaw_dev

  —

  本地跑 Coder 最大的价值是数据不出域，代码仓库丢给云端 API 始终不放心，Apache 2.0 商用随便改，这比 Claude 香多了。

---

• 

  BitSharkSimmons

  —

  说实话跟 Qwen3-Coder-Next 比提升没想象中大，3.5-Coder 7B 在 SWE-bench 上可能就多了两三个点，但模型体积涨了不少，部署性价比存疑。

---

• 

  DAram

  —

  试了 Coder 7B 做 Python 代码补全，日常够用，但遇到复杂的数据结构题就露馅了，还是要靠 32B。

---

• 

  Sophia_SimmonsII666

  —

  最香的是阿里云百炼 API 价格，Qwen3.5-Coder 的 Plus 版本才 4 块多每百万输出 token，比 Claude 便宜 10 倍，做批处理完全不心疼。但注意长上下文阶梯计价，超过 128K 输入价格直接翻倍，大项目要算好成本，别等账单来了才傻眼。

---

• 

  Evelyn_Miller_Pro

  —

  部署踩了个大坑，官网写的 HF 模型卡跟我下载的不一致，折腾了一上午才跑起来，新手建议直接用 Ollama 官方版。

---

• 

  PilarLozano

  —

  用 openclaw 接 Coder 做了个自动修 Bug 的 Agent，昨晚跑了 4 个小时没中断，GitHub issue 闭环率大概六成左右，偶尔会改坏。不过对于简单的语法错误修复，成功率接近九成，日常维护够用了。

---

• 

  JaniceHall_Pro82

  —

  yyds，32B 量化跑 llama.cpp 写 Rust 的网络服务，自己处理 API 轮询加数据库写入，全程没插手。

---

• 

  JamesPatelX

  —

  代码审查功能确实不错，帮我揪出了一个多线程竞态条件，以前这种 Bug 至少得 debug 半天。

---

• 

  BettyFloresSr

  —

  问一下，7B 版本的 GGUF 文件大概多大？想在 MacBook 上跑，不知道 16GB 内存够不够。

---

• 

  Debra.Reed_2023

  —

  7B 版大概 4-5GB 量化文件，16GB 内存完全够。M 芯片还能用 Metal 加速，我 M2 跑大概 20t/s，日常补全很流畅。

---

• 

  ixtyoucl5

  —

  跟 DeepSeek-V3.2 比，Coder 在中文注释理解上强很多，DeepSeek 经常把中文注释写得像机翻。

---

• 

  HsshBase

  —

  Apache 2.0 许可就是定心丸，公司法务审查后直接放行，不用走额外的合规流程。

---

• 

  Frank.Cook_7

  —

  遇到个问题，长上下文推理到 100K 左右的时候，质量明显下降，指令跟随开始漂移，不知道是不是我量化精度选太低了。用的是 Q4_K_M 量化，有没有大佬试过 Q8 甚至 FP16 在高上下文长度下的表现，会不会好一些。

---

• 

  HelenSchmidt

  —

  回不去了，用惯本地模型之后再也忍受不了云端 API 的延迟和配额限制，哪怕慢一点也是自己的。

---

• 

  BobbyGreen_2022

  —

  发现 Coder 32B 的 FIM 内联补全能力比同参数的通用模型好不少，特别是写 SQL 的时候效率翻倍。

---

• 

  PRamirez_2020

  —

  7B 跑 Agent 多步骤任务容易崩，三步以上就开始丢上下文，建议至少上 32B 或者用 API 版。不过如果只是做单步代码补全或者写单元测试，7B 的性价比反而是最高的。

No reviews yet