# GLM 5.2 vs Kimi K2.7 Code: The Definitive Guide for Coding

These are two open-weight models released in June 2026 just one day apart, both Mixture-of-Experts systems and both aimed at developers but under that surface similarity sit two very different design choices.

**GLM 5.2** — launched on June 13, 2026 by Z.ai — is a 744B-parameter model with about 40B active parameters per token, Its core purpose is long-horizon coding on extremely large contexts. The 1 million token window is central to the model design, enabled by sparse-attention optimizations that reduce FLOPs per token at maximum context. It is released under the [MIT license](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md) with [weights on Hugging Face.](https://huggingface.co/zai-org/GLM-5.2)

**Kimi K2.7 Code** — released on June 12, 2026 by Moonshot AI — is a roughly 1 trillion parameter model with 32B active parameters per token and a MoE architecture using hundreds of routed experts with 8 active per token. It has a 256K context window, requires reasoning mode, and a HighSpeed variant that reaches up to 260 tokens per second was announced a few days later. It is natively multimodal, supporting visual input through an integrated vision encoder. [Its license is Modified MIT](https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/LICENSE) and [weights are here on Hugging face](https://huggingface.co/moonshotai/Kimi-K2.7-Code).

## Table of contents

1. [Benchmark](#benchmark)
2. [Charts](#Charts)
3. [Coding & code review](#coding)
4. [Long Document](#long-document)
5. [Batch Extraction](#batch-extraction)
6. [Creative Writing](#creative-writing)
7. [Image Understanding / OCR](#image-understanding)
8. [Latency & Speed](#latency-speed)
9. [When to choose GLM 5.2 — and when to choose Kimi K2.7](#when)
10. [FAQ](#faq)

---

# Benchmark

## Capabilities score

| Benchmark | GLM 5.2 | Kimi K2.7 Code | Notes |
|---|---|---|---|
| SWE-bench Pro | **62.1%** | ~58.6% from K2.6 | K2.7 still lacks broad independent scoring |
| Terminal-Bench 2.1 | **81.0%** | 71.0% estimate |  |
| FrontierSWE | **74.4%** | N/A |  |
| Kimi Code Bench v2 | N/A | **62.0%** | Internal benchmark |
| Program Bench | N/A | **53.6%** | Internal benchmark |
| MLS Bench Lite | N/A | **35.1%** | Reported as around GPT-5.5 level |
| MCP Mark Verified | N/A | **81.1%** | Above Claude Opus 4.8 in Moonshot’s reported comparison |
| HumanEval | ~85.2% | ~94.2% | Third-party estimates |
| GPQA-Diamond | **91.2%** | N/A |  |
| AIME 2026 | **99.2%** | N/A |  |

To compare models with uneven benchmark coverage — GLM 5.2 has public scores on common international benchmarks, while Kimi K2.7 still relies heavily on proprietary ones — a normalized weighted score is more useful than a raw leaderboard view

| Model | Capabilities score (0–100) | Basis |
|---|---|---|
| GLM 5.2 | **78/100** | SWE-bench Pro, Terminal-Bench, FrontierSWE, GPQA, AIME |
| Kimi K2.7 Code | **70/100** | Kimi Code Bench, MCP Mark, MLS Bench, HumanEval estimate |

That gap partly reflects evidence quality, not just model quality. Kimi K2.7 may improve once more independent benchmarks are published, but as of June 18, 2026 GLM 5.2 is easier to verify from public data.

### Pricing efficiency

Kimi K2.7 is cheaper on both input and output. GLM 5.2 becomes more competitive when long repeated prompts benefit from caching, especially on large-repository workflows.

| Metric | GLM 5.2 | Kimi K2.7 Code |
|---|---|---|
| Input | $1.40 / 1M tokens | **$0.95 / 1M tokens** |
| Output | $4.40 / 1M tokens | **$4.00 / 1M tokens** |
| Cache hit | about $0.26 / 1M tokens | **$0.19 / 1M tokens** |

### Context window size

This is the clearest difference in the whole comparison. GLM 5.2 can keep entire large repositories or very large document sets inside a single prompt in ways Kimi K2.7 cannot.

| Model | Maximum context | Operational note |
|---|---|---|
| GLM 5.2 | **1,000,000 tokens** | 1M mode is available through the dedicated variant |
| Kimi K2.7 Code | 262,144 tokens | Recall reportedly weakens past about 180K |

### Output capacity

GLM 5.2 explicitly documents its output ceiling. Moonshot has not yet published a precise maximum output figure for Kimi K2.7 Code.

| Model | Maximum output |
|---|---|
| GLM 5.2 | **131,072 tokens** |
| Kimi K2.7 Code | Not officially published |

### Recency

The freshness difference is basically zero. Both were among the most important open-weight releases of mid-June 2026.

| Model | Release date | License |
|---|---|---|
| GLM 5.2 | June 13, 2026 | MIT |
| Kimi K2.7 Code | June 12, 2026 | Modified MIT |

## Global composite score

The composite score combines benchmark strength, pricing, context window, recency, versatility, and output capacity.

| Dimension | Weight | GLM 5.2 | Kimi K2.7 Code |
|---|---|---|---|
| Benchmark performance | 30% | **9/10** | 7/10 |
| Pricing efficiency | 20% | 7/10 | **9/10** |
| Context window | 20% | **10/10** | 5/10 |
| Recency | 5% | 10/10 | 10/10 |
| Versatility | 15% | 6/10 | **9/10** |
| Output capacity | 10% | **9/10** | 6/10 |
| **Composite score** | **100%** | **8.15 / 10** | **7.55 / 10** |

GLM 5.2 wins the overall composite mainly because of its massive context window and stronger public benchmark trail. Kimi K2.7 closes the gap with lower pricing and native multimodality.

---

# Charts

## Benchmark Scores: **how “good” each model is at coding and reasoning tests**.

Each group of two bars = one benchmark.

- **Blue (GLM 5.2)** vs **pink (Kimi 2.7)**.
- The **higher bar** means **better performance** on that test.

**How to use this chart:**

- For **standard coding tasks** (bug fixing, implementing specs), look at bars like **SWE-bench Pro / Terminal-Bench**: if GLM’s bar is clearly higher, it’s the safer choice for complex code changes.
- Where you see **“N/A”**, it just means “no reliable public data yet”, not that the model is bad; don’t over‑penalize, but treat it as unknown.

![](https://regolo.ai/wp-content/uploads/2026/06/benchmark_scores-1024x683.png)Think of it as: *“If I throw a hard coding problem at both, which one statistically solves more of them?”* — that’s what the taller bars are telling you.

## Context & Output chart: **how much code / text you can fit into one request** and **how long the answer can be**.

![](https://regolo.ai/wp-content/uploads/2026/06/context_output-1-1024x683.png)In short: this chart tells you *“Can this model see my entire problem at once, or do I need to chunk it?”*

For each model you see:

- **Context Window (K tokens)**: how much you can send in.
    GLM ~1000K (1M), Kimi ~262K.
- **Max Output (K tokens)**: how much it can reply with in one shot.
    GLM has a clear bar; Kimi is “N/A / undisclosed”.

How to use it:

- If you want to paste in **huge repos, massive logs or multiple big docs at once**, GLM’s huge context makes a practical difference.
- If your typical task is **single files or small services**, Kimi’s smaller window is usually enough.
- For **very long generated artifacts** (big migrations, long reports), GLM’s high output bar means fewer “continue” calls.

## Pricing chart: **“If I spam these models with tokens, who kills my budget faster?”**

Three groups on the x‑axis:

- **Input** = what you send,
- **Output** = what you get back,
- **Cache hit** = repeated context that gets billed cheaper.

Lower bar = **cheaper**.

How to use it:

- If your workflow does **lots of small calls** (Copilot‑style helper, frequent refactors), the **Input** bar matters a lot.
- If you generate **long answers** (scaffolding whole files, large reviews), look at **Output**.
- If you **reuse the same big prompt** (same repo context, many similar queries), the **Cache Hit** bar tells you which one stays cheaper in the long run.

![](https://regolo.ai/wp-content/uploads/2026/06/pricing-1-1024x683.png)When money matters more than 1–2 extra percentage points of quality, pick the model with the **shorter bars** for the pattern you use most.

## Speed chart: **how fast the models spit out tokens** under different modes.

![](https://regolo.ai/wp-content/uploads/2026/06/speed-1-1024x683.png)So you can read this chart as *“Which option feels less like waiting on a slow CI job?”*

Each bar is a variant:

- GLM average / GLM best provider,
- Kimi standard / Kimi HighSpeed.

Higher bar = **more tokens per second** → feels more responsive in your editor.

How to use it:

- For **interactive coding** (chat in IDE, pair‑programming feel), the **fastest bars** matter: Kimi HighSpeed will *feel* snappier.
- For **batch scripts / offline jobs**, speed still matters, but you might prioritize quality or context instead.
- If you hate “waiting for the model to finish scrolling”, pick the mode with the tallest bar that still meets your quality needs.

## Composite Radar chart: it's a summary map of trade-offs GLM 5.2 vs Kimi 2.7 Code

The further the colored shape goes toward the edge on a spoke, the **better that model is on that dimension**:

- Blue shape = GLM 5.2
- Pink shape = Kimi 2.7

How to use it:

- If your work is **repo‑heavy backend dev**, look at **Context / Output / Benchmark**: GLM’s shape sticks out more there → better for deep, long tasks.
- If you do **tool‑heavy, multimodal, UI + API hacking**, look at **Pricing / Versatility**: Kimi’s shape is larger → cheaper and more flexible (images, agents).
- Use it to pick a **default model per project type**: e.g. “infra monorepo → GLM”, “agentic coding with screenshots → Kimi”.

![](https://regolo.ai/wp-content/uploads/2026/06/composite_radar-1024x683.png)You can mentally map it to: *“Person A is stronger, Person B is faster and more flexible — who do I put on which ticket?”*

---

# Coding & code review

GLM 5.2 is the top open-weight model on SWE-bench Pro at 62.1% and also reports 81.0% on Terminal-Bench 2.1. On FrontierSWE it reaches 74.4%, placing i**t very close to leading proprietary systems** in that benchmark family.

Kimi K2.7 Code **looks especially strong in agent-heavy coding workflows rather than traditional benchmark transparency**. It reports 81.1% on MCP Mark Verified and 62.0% on Kimi Code Bench v2, up 21.8% from its predecessor on that internal benchmark.

The main weakness is evidence visibility: broad independent SWE-bench Pro or Terminal-Bench results for K2.7 were not yet widely available at the time of writing.

For single-file code review, fast debugging passes, or interactive assistant behavior, **Kimi K2.7 Code** is often the more efficient pick. For repo-wide review, multi-file patches, and complex long-horizon engineering tasks, **GLM 5.2** is the better choice.

---

# Long document analysis

This is where GLM 5.2 stands out most clearly, one million tokens is enough for very large repositories, long production logs, or large bundles of contracts in a single pass. Z.ai’s ecosystem documentation explicitly frames the model around repository-scale and long-context reasoning use cases.

Kimi K2.7’s 256K context is still large enough for single big documents, medium repositories, and most standard technical analysis tasks, but reported recall degradation beyond roughly 180K makes it a weaker fit once prompt size becomes extreme.

**For anything above about 200K tokens in one request, GLM 5.2 is the safer choice.**

---

# Batch extraction

This "feature" depends on large context, long output and low cost.

On context and maximum output, GLM 5.2 leads by a wide margin, on raw price, Kimi K2.7 is better – for repeated extraction templates over many similar files, Kimi’s lower cache-hit pricing makes it especially attractive.

So the split is simple: for many documents that each stay under about 200K tokens, **Kimi K2.7 Code** usually gives better economics, instead for oversized files or one-shot extraction on very large contexts, **GLM 5.2** is the more capable option.

---

# Creative writing

Neither model was primarily built as a creative-writing system. Both are engineering-first products.

GLM 5.2 benefits from long-context consistency and very large output capacity, which matters for long-form narrative continuity, It is not generally described as unusually imaginative, but it is better suited to holding long threads together over large outputs.

Kimi’s broader K2 family has been described as capable in long-form and multiturn writing, and K2.7 inherits that base, but its mandatory reasoning mode adds overhead even on simple creative tasks.

**For short iterative creative work, the HighSpeed variant makes Kimi more appealing; for long, coherence-heavy writing, GLM 5.2 has the stronger structural advantage.**

---

# Image understanding / OCR

**GLM 5.2 is text-only:** It does not take image input If a workflow includes scanned PDFs, screenshots, diagrams, or visual layouts, a separate OCR or vision stage is required. In Z.ai’s own stack, that role is handled by **GLM-OCR**, a separate compact multimodal OCR model built for document understanding.

**Kimi K2.7 Code is natively multimodal:** It supports image and video input through its integrated visual encoder that makes it the natural choice for coding workflows involving UI screenshots, diagrams, visual debugging, or OCR-like reasoning inside one model loop.

**For any use case involving images, video, or scanned documents, Kimi K2.7 Code is the clear winner.**

---

# Latency & speed

| Metric | GLM 5.2 | Kimi K2.7 Code |
|---|---|---|
| Output speed | 59–166 tok/s depending on provider | 52 tok/s standard, 180–260 tok/s HighSpeed |
| Time to first token | roughly 1.0–14.8s depending on provider | around 2.3s |
| Local use | possible with quantized setups | generally not practical on ordinary single-node hardware |

GLM 5.2 shows large provider variance in the cloud. Kimi K2.7 HighSpeed is the more predictable choice for real-time agent workflows where snappy interaction matters.

For local deployment, GLM 5.2 is far more realistic because open weights and quantized paths are already discussed in public setup guides. Kimi K2.7’s scale makes practical self-hosting far harder for most developers.

---

# When to choose GLM 5.2 — and when to choose Kimi K2.7

Choose **GLM 5.2** if your codebase exceeds about 200K tokens, if you need repo-scale refactoring, if long-horizon engineering matters, or if you want very long outputs in one response. It is also the better pick when you want self-hosting flexibility and your workflow is entirely text-based.

Choose **Kimi K2.7 Code** if your workflow includes screenshots, diagrams, scanned documents, or other visual inputs. It is also the better fit for tool-heavy agent loops, lower-cost batch work, and real-time interactive coding assistants where latency matters. As long as your working context stays below roughly 180K–256K tokens, it is often the more economical day-to-day option.

---

# FAQ

**Does GLM 5.2 support images or PDFs?**
No. GLM 5.2 is text-only. For OCR or visual document analysis, Z.ai provides GLM-OCR as a separate model, while Kimi K2.7 accepts images and video natively.

**Can Kimi K2.7 Code really handle a 1M-token codebase?**
No. Its maximum context is about 262K tokens, and practical recall reportedly weakens beyond around 180K. GLM 5.2 is the one that reaches 1M tokens in a single request path.

**What is the effective cost for a code review task over 50 files, around 30K input tokens and 5K output tokens?**
At published rates, GLM 5.2 works out to about $0.042 for input plus about $0.022 for output, or roughly $0.064 total. Kimi K2.7 works out to about $0.0285 for input plus about $0.020 for output, or roughly $0.0485 total. That is about a 24%–25% savings in Kimi’s favor for that workload.

**Can Kimi K2.7 Code be used without reasoning mode?**
No. Public documentation describes the thinking mode as required rather than optional.

**Does GLM 5.2 depend on NVIDIA-only training?**
No. Public descriptions note Z.ai’s training lineage includes non-NVIDIA hardware in its broader ecosystem, while inference availability depends on the provider.

---

Sta**rt your free 30-day trial at [regolo.ai](https://regolo.ai/) and deploy LLMs with complete privacy by design.**

👉 [Talk with our Engineers](https://regolo.ai/contacts/) or [Start your 30 days free →](https://regolo.ai/pricing)

---

- [Discord](https://discord.gg/ZzZvuR2y) - Share your thoughts
- [GitHub Repo](https://github.com/regolo-ai/) - Code of blog articles ready to start
- Follow Us on X [@regolo\_ai](https://x.com/regolo_ai)
- Open discussion on our [Subreddit Community](https://www.reddit.com/r/regolo_ai/)
- Full list of model available: [Models](/models)

---

*Built with ❤️ by the Regolo team. Questions? [regolo.ai/contact](https://regolo.ai/contact)* or chat with us on [Discord](https://discord.gg/ZzZvuR2y)