Open-weight models for coding.

Shortlist of open AI models for code review, refactoring, and agent loops — compare specs and test them with real prompts.

Choosing an open-weight AI model for software engineering is about matching your specific task to three factors: context window size, reasoning depth, and turnaround speed. Rather than relying on synthetic benchmarks, this curated shortlist focuses on models suitable for everyday development, bug triage, refactoring, and agent tool calling.

Every model below is free to test on UsingOpen within your daily prompt limit. To evaluate two candidates simultaneously, click Compare mode to inspect both answers side by side.

Coding models shortlist (12 models)

Open-weight models on UsingOpen with maker, parameters, context window, price and capabilities
ModelMakerParamsContextPrice / MTokImagesToolsZDR
GLM 5.3 FlashZ.ai320B1M (750,000 words)$0.25YesYesYesChat free
Qwen3.8 FlashAlibaba—1M (750,000 words)$0.47YesYesNoChat free
Qwen3.8 27BAlibaba27B262k (200,000 words)$3.20YesYesYesChat free
DeepSeek V4 FlashDeepSeek284B1M (750,000 words)$0.18NoYesYesChat free
DeepSeek V4 ProDeepSeek1.7T1M (750,000 words)$1.98NoYesYesChat free
GPT-OSS 120BOpenAI117B128k (95,000 words)$0.60NoYesYesChat free
GLM 5.3Z.ai753B1M (750,000 words)$2.43NoYesYesChat free
InklingThinking Machines975B1M (750,000 words)$4.05YesYesYesChat free
Kimi K3Moonshot AI2.8T1M (750,000 words)$8.50YesYesYesChat free
DeepSeek V4 Pro 0423DeepSeek1.6T1M (750,000 words)$2.09NoYesYesChat free
Qwen3.8 2.4TAlibaba2.4T1M (750,000 words)$6.00NoYesYesChat free
MiniMax M3MiniMax427B1M (750,000 words)$1.20YesYesYesChat free

How to shortlist a coding model

  • High-volume assistance & quick iterations: Models like DeepSeek V4 Flash and Qwen3.8 Flash activate a small subset of parameters per token, keeping responses rapid while retaining a full 1-million-token context window.
  • Deep architectural reasoning & complex refactoring: When you need careful multi-step verification and fewer hallucinations on tricky logic, DeepSeek V4 Pro 0423, GLM 5.3, or Qwen3.8 2.4T deliberate more thoroughly before answering.
  • Tool-calling & agent loops: Models with native tool support, such as GPT-OSS 120B and MiniMax M3, are optimized for structured function calling and step-by-step agentic execution.

Web chat vs local desktop execution

UsingOpen Web Chat limitations: The free web chat offers 10 prompts per user per day with streaming generation. It is designed for conversational testing and snippet reviews; repository indexing, file attachments, and image uploads are not currently supported in the web composer.

Local desktop work with UsingOpen Studio: For commercial codebases, repository-level analysis, offline execution, and zero data exposure, consider UsingOpen Studio. It runs open weights natively on your Mac, Windows, or Linux hardware with no per-token fees, no prompt limits, and no cloud dependency.

How to evaluate these models on your own code

Leaderboard scores rarely reflect how an LLM handles your codebase's conventions. Instead of guessing, run shortlisted models against this reproducible evaluation prompt in Compare mode:

Review this code change.

1. Identify concrete correctness bugs.
2. Separate confirmed bugs from possible risks.
3. Give a minimal failing example for each confirmed bug.
4. Suggest the smallest safe correction.
5. State what you could not verify.

Popular head-to-head coding comparisons

Frequently asked questions

Which open-weight model is best for coding?

For high-volume coding help and fast script generation, DeepSeek V4 Flash and Qwen3.8 Flash offer 1-million-token context windows at very low token costs. For deep architectural reasoning, difficult refactoring or complex bug-hunting, flagships like DeepSeek V4 Pro, Qwen3.8 2.4T, and GLM 5.3 provide more deliberate step-by-step thinking.

Does context window size matter for programming tasks?

Yes. A 128k–256k window is plenty for single functions, pull requests or small modules. However, repository-wide analysis, multi-file refactoring, and extensive log traces require 1-million-token windows (roughly 750,000 words) so the model can inspect all interdependent files in one conversation.

How should I choose between fast Flash models and larger flagships?

Start with a fast, cost-effective model like DeepSeek V4 Flash or Qwen3.8 Flash. If the model fails to follow instructions on edge cases or produces incorrect logic, escalate the task to a reasoning model like DeepSeek V4 Pro or Qwen3.8 2.4T. You can test both side by side on UsingOpen in Compare mode.

Can I run these coding models locally with complete privacy?

Yes. While UsingOpen offers free hosted chat (up to 10 prompts per day), our desktop application, UsingOpen Studio, runs open models 100% locally on your own Mac, Windows, or Linux computer without sending your code to any cloud server.

Does the UsingOpen web chat support direct repository uploads?

Not yet. The free web chat composer supports text and code snippets up to the character limit. To index entire local Git repositories and run offline coding agents without prompt limits, use UsingOpen Studio.