Kimi K2.7 Code: 1T Open-Source Coding Model Released

Moonshot AI released Kimi K2.7 Code on June 18, 2026. The coding model carries 1 trillion parameters and a 128,000-token context window. It ships as open weights under a modified Apache 2.0 license. The company announced the launch on its official homepage and published model cards on Hugging Face. Developers can download the weights without a paid key or approval queue. The model targets code generation, debugging, test writing, and repository-scale edits. Moonshot AI says the model trained on 18 trillion tokens with heavy code, math, and reasoning data. This release is part of a rapid open-source push covered in today’s AI updates. More details are on the Moonshot AI homepage.

Moonshot AI built the model on the Kimi K2 architecture that already powers its consumer assistant. This release is a separate coding-focused checkpoint rather than a general chat model. The weights come from the vendor announcement page, not a silent drop. Moonshot AI published benchmark tables and a technical report linked from its homepage. The company says Kimi K2.7 Code reaches 72.4 percent on SWE-bench Verified and 68.1 percent on LiveCodeBench. Those results put it ahead of DeepSeek V4 and Qwen3-Coder in vendor-run tests. The open license means developers can fine-tune, distill, and self-host it without usage-based fees like the ones hitting GitHub Copilot. That shift matters for budget-conscious teams. Independent tests are still pending across several leaderboards.

Why this matters is straightforward. Closed coding models from OpenAI and Anthropic still lead on some agentic benchmarks, but their API pricing has become less predictable for free and low-cost tiers. Moonshot AI is giving away a 1T-parameter model that can run on a single high-end GPU with quantization. That changes the math for startups that cannot afford per-token markups. The code-specific design also narrows the gap with proprietary tools. In early tests cited by the vendor, Kimi K2.7 Code beats GPT-5.1 Codex on HumanEval and matches Claude Opus 4.8 on SWE-bench Verified. Those comparisons come from Moonshot AI, so independent verification is still needed. The release arrives as several providers tighten free API access, detailed in AI free tier limits.

Moonshot AI released the model on a Thursday, a common day for open model drops. The timing matters because developer patience with paid coding tools has worn thin. GitHub Copilot and Cursor both introduced usage-based billing changes in June 2026, which prompted backlash. An open 1T model offers a fallback for teams that want predictable local inference. It also gives researchers a base model for distillation. Many developer teams are now evaluating self-hosted models as a cost hedge. Kimi K2.7 Code is not the only open coding launch this month, but its size and benchmark claims make it one of the most significant this quarter. The launch page includes example commands and a model card.

How Do the Top Options Compare?

Benchmark scores reflect vendor-reported or leaderboard results as of June 2026. Real-world performance depends on quantization and hardware.

1. Kimi K2.7 Code , Local open-source coding with no per-token fees

Kimi K2.7 Code is the main event. The model uses a 1-trillion-parameter mixture-of-experts architecture with 8 active parameters per token. It handles 128,000 tokens of context, enough for entire codebases or large documentation files. Moonshot AI released the weights under a modified Apache 2.0 license that allows commercial use, fine-tuning, and distillation. The company asks for attribution in derivative models but does not charge royalties. You can find the official release note on the Moonshot AI homepage. The model is not hidden behind a proprietary API.

Running it locally is feasible with aggressive quantization. In FP8 form it needs about 520GB of VRAM across eight 80GB GPUs. A 4-bit GPTQ version drops that to roughly 300GB, which fits on four RTX 4090s or two H100s. Moonshot AI also published an MLX version for Apple silicon, though the 1T-parameter size makes it slow on anything below 192GB of unified memory. The launch page includes example commands for vLLM, SGLang, and text-generation-webui. If you want a simpler free route first, compare options in best free AI models 2026.

On benchmarks, Kimi K2.7 Code leads the open coding pack. Vendor results show 72.4 percent on SWE-bench Verified, 68.1 percent on LiveCodeBench, and 96.8 percent on HumanEval. These numbers are strong but not independently reproduced as of launch day. On Aider Polyglot the model scores 67.5 percent, below GPT-5.1 Codex but above Claude Opus 4.8. The model also supports tool calling, fill-in-the-middle editing, and diff generation. Those features make it a direct rival to paid coding assistants.

For production use, teams should start with the 4-bit quantized version. Memory bandwidth matters more than raw parameter count at inference time. A single H100 with 80GB cannot serve the full model without model parallelism. Four 80GB GPUs work well for a small team. The model outputs code with strong indent style and docstrings. It is not perfect on long agent loops. But for code review, unit test generation, and refactoring, it is a serious open option.

Key strengths:

• ✅ Strong benchmark scores on SWE-bench Verified and HumanEval for an open model.
• ✅ Modified Apache 2.0 license allows commercial use and fine-tuning without fees.
• ✅ 128K context handles large codebases and long debugging sessions.
• ✅ Multiple quantization options let it run on 4 to 8 GPUs.
• ❌ Requires significant VRAM even with 4-bit quantization.
• ❌ Vendor benchmarks are not yet independently verified.
• ❌ Apple silicon support is slow for the full model.

Who it’s for: Teams that want a self-hosted coding model with no per-token cost and have access to high-end GPUs.

2. DeepSeek V4 , Budget API coding with a larger open context window

DeepSeek V4 is the closest open competitor in parameter count. It packs 1.6 trillion parameters with 16 active per token and a 256,000-token context window. The model launched earlier in June 2026 under an MIT license, which is even more permissive than Moonshot AI’s terms. You can grab weights from the DeepSeek homepage. It scores 70.8 percent on SWE-bench Verified and 72.3 percent on LiveCodeBench in vendor tests. Those results are close to Kimi K2.7 Code but with a much larger context window.

DeepSeek V4 shines for API budgeting. The company offers an open-weight model plus a hosted endpoint at low per-token rates. This gives developers the option to start on the API and later self-host without changing the model family. The larger context helps with long repositories, logs, and multi-file refactors. However, the 1.6T parameter count makes local deployment harder than Kimi K2.7 Code. Even with 4-bit quantization, DeepSeek V4 needs about 480GB of VRAM. Most teams will rent cloud GPUs instead. The major AI API pricing updates June 2026 article shows why predictable open API pricing is now a selling point.

On code-specific tasks, DeepSeek V4 excels at SQL, Python, and JavaScript. It also has stronger multilingual code documentation than the Kimi model, according to early community tests. But it lacks a dedicated fill-in-the-middle mode in the current release. That makes it slightly less useful for IDE integration. Still, the MIT license and huge context make it a solid alternative. Teams already using DeepSeek’s chat API can switch to V4 coding endpoints without a new vendor.

The main downside is hardware cost. A full 1.6T model requires a cluster for smooth inference. Quantization helps but still leaves you with a model that needs four to eight high-end GPUs. For a small team, Kimi K2.7 Code is often easier to self-host. For an enterprise that already rents H100s, DeepSeek V4’s larger context may justify the extra overhead. The choice usually comes down to whether you need 256K tokens or a smaller memory footprint.

Key strengths:

• ✅ MIT license is fully permissive for commercial use and redistribution.
• ✅ 256K context window handles very large repositories and logs.
• ✅ Low-cost hosted API offers an easy starting point before self-hosting.
• ✅ Strong multilingual code documentation and reasoning.
• ❌ 1.6T parameters make local deployment expensive.
• ❌ No dedicated fill-in-the-middle mode for IDE autocomplete.
• ❌ SWE-bench Verified score slightly below Kimi K2.7 Code.

Who it’s for: Developers who want a permissive open license and a large context window, and who may start on a cheap hosted API.

3. Qwen3-Coder , Multilingual coding and smaller GPU footprints

Qwen3-Coder is Alibaba’s open coding model from the Qwen family. It uses 480 billion parameters with 12 active per token and a 256,000-token context window. The model is available under Apache 2.0. You can find the release details on the Alibaba Cloud site. In vendor benchmarks, Qwen3-Coder scores 69.2 percent on SWE-bench Verified and 71.6 percent on LiveCodeBench. It falls behind Kimi K2.7 Code on English coding tasks but often leads on Chinese, Japanese, and Korean documentation.

The smaller parameter count is a real advantage. Qwen3-Coder runs with 4-bit quantization in about 120GB of VRAM. That means a single workstation with two RTX 6000 Ada cards can serve it. The model supports fill-in-the-middle, diff generation, and tool calling. It also has a long context that fits most monorepos. For smaller teams that cannot afford eight H100s, Qwen3-Coder is often the practical choice. The model works well with vLLM and llama.cpp.

Qwen3-Coder is not as strong on complex agentic coding tasks as the 1T-class models. On Aider Polyglot it scores 63.2 percent, which is respectable but not class-leading. It also struggles with very low-level C++ and Rust code compared to Kimi K2.7 Code. But for common web, mobile, and data-science code, it is reliable. The Apache 2.0 license allows commercial products without attribution requirements. That makes it an easy drop-in for companies building private coding tools.

The main limitation is peak capability. If you need the highest SWE-bench score, Kimi K2.7 Code or DeepSeek V4 will serve better. But many teams do not need the top score. They need a model that fits their existing hardware and license review. Qwen3-Coder hits that balance. It is the kind of model that can quietly power an internal code review bot without a large GPU bill.

Key strengths:

• ✅ 480B size runs on two high-end workstation GPUs with quantization.
• ✅ Apache 2.0 license has no attribution or commercial restrictions.
• ✅ Strong multilingual coding, especially for CJK languages.
• ✅ Long 256K context fits most codebases.
• ❌ Lower SWE-bench Verified score than 1T open models.
• ❌ Weaker on low-level C++ and Rust than Kimi K2.7 Code.
• ❌ Less community tooling for MLX and Apple silicon.

Who it’s for: Teams that need a capable open coder without a large GPU cluster and want painless commercial licensing.

4. Zyphra Zaya , Edge and low-resource coding

Zyphra Zaya is a small open-weight reasoning model designed for edge devices. It has 7 billion parameters and a 32,000-token context window. The model ships under Apache 2.0. Zyphra released it in late May 2026 with a focus on on-device coding and reasoning. You can read the full release notes on the Zyphra homepage. Zaya is not a direct Kimi K2.7 Code competitor in raw capability. It scores 54.1 percent on SWE-bench Verified and 48.7 percent on LiveCodeBench. But those scores are impressive for a 7B model that runs on a laptop.

Zaya’s practical advantage is hardware fit. The model runs in 4-bit quantization with under 8GB of VRAM. It can execute on a MacBook Pro, a high-end Android phone, or a Raspberry Pi with acceleration. That makes it useful for local autocomplete, code linting, and simple refactors without sending code to the cloud. After free coding assistants like Cursor and Zed tightened limits, a local 7B model regained attention. Zyphra also published a specialized 2B variant for smaller devices. The small size means it does not need a GPU cluster or a paid API.

Zaya cannot handle large repository-scale edits or complex algorithmic problems. Its 32K context limits it to single files or small modules. For big jobs, teams will still want Kimi K2.7 Code, DeepSeek V4, or Qwen3-Coder. But for privacy-sensitive coding and offline development, Zaya fills a distinct role. It also works as a benchmark for how far edge coding has come. The model can run entirely offline, which matters in regulated sectors.

The tradeoff is clear. You get absolute local control but limited capability. Zaya is not a replacement for a 1T coding model. It is a complementary tool for quick, private tasks. If you need a full repository agent, look at the larger open models. If you need a small autocomplete engine that never phones home, Zaya is a solid start.

Key strengths:

• ✅ Runs locally on a laptop or phone with under 8GB VRAM.
• ✅ Apache 2.0 license allows unrestricted commercial use.
• ✅ Good privacy for offline coding and code linting.
• ✅ 2B variant available for even smaller devices.
• ❌ 32K context cannot handle large codebases.
• ❌ Benchmark scores are far below 1T and 480B coding models.
• ❌ Limited tooling for integrated development environments.

Who it’s for: Privacy-focused developers who need a lightweight local coder for single-file tasks without cloud dependencies.

Frequently Asked Questions

What Should You Remember?

• Kimi K2.7 Code is a 1T-parameter open coding model released on June 18, 2026.
• Modified Apache 2.0 license allows commercial use, fine-tuning, and distillation without royalties.
• Benchmarks put it at 72.4 percent on SWE-bench Verified and 68.1 percent on LiveCodeBench.
• Local deployment is possible with 4-bit quantization but still needs about 300GB of VRAM.
• DeepSeek V4 offers a larger context window and MIT license but is heavier to self-host.
• Qwen3-Coder is the practical choice for teams with one or two high-end GPUs.
• Zyphra Zaya handles offline single-file coding on edge devices under 8GB VRAM.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.