Meta Description: Four major Chinese AI labs dropped open-weight models within a 12-day window in April–May 2026. This guide breaks down DeepSeek V4, Kimi K2.6, GLM-5.1, and MiniMax M2.7 — specs, benchmarks, agentic capabilities, and inference costs compared.

—

Table of Contents

1. [Why April 2026 Marked a Turning Point](#1-why-april-2026-marked-a-turning-point)
2. [The 12-Day Release Window](#2-the-12-day-release-window)
3. [DeepSeek V4: The Efficiency Revolution](#3-deepseek-v4-the-efficiency-revolution)
4. [Kimi K2.6: The Coding & Agent Swarm Champion](#4-kimi-k26-the-coding–agent-swarm-champion)
5. [GLM-5.1: Agentic Engineering Powerhouse](#5-glm-51-agentic-engineering-powerhouse)
6. [MiniMax M2.7: Self-Evolving Open Model](#6-minimax-m27-self-evolving-open-model)
7. [Head-to-Head Benchmark Comparison](#7-head-to-head-benchmark-comparison)
8. [Inference Cost: Why Western Models Can’t Compete](#8-inference-cost-why-western-models-cant-compete)
9. [Who Should Use Which Model?](#9-who-should-use-which-model)
10. [The Bottom Line](#10-the-bottom-line)

—

1. Why April 2026 Marked a Turning Point

For years, the open-weight AI landscape was dominated by Western labs — Meta’s Llama series, Google’s Gemma line, and Mistral’s models kept pulling ahead in capability. That narrative shattered in April 2026.

Between April 12 and April 24, 2026, four of China’s most advanced AI labs each released open-weight flagship models, collectively covering reasoning, coding, agent orchestration, and self-evolution. The timing was deliberate: each lab timed their release to overshadow the others, creating what analysts called the “Chinese AI open-weight blitz.”

The result? A developer ecosystem suddenly had access to frontier-class open models that rival GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro — at a fraction of the cost.

—

2. The 12-Day Release Window

Here’s the exact timeline that sent shockwaves through the AI community:

| Date | Lab | Model | Highlight |
|——|—–|——-|———–|
| April 12, 2026 | MiniMax | M2.7 | Open weights released, self-evolving agent model |
| April 20, 2026 | Moonshot AI | Kimi K2.6 | #1 open-weight model, Agent Swarm, coding SOTA |
| April 24, 2026 | DeepSeek | V4 Pro & V4 Flash | Two-tier release, 27% inference FLOPs vs V3.2 |
| Early May 2026 | Z.ai | GLM-5.1 | SOTA SWE-Bench Pro (58.4%), agentic engineering |

Four drops in 12 days. No coincidence — each release was timed to dominate headlines, benchmark leaderboards, and developer attention.

—

3. DeepSeek V4: The Efficiency Revolution

Release Date: April 24, 2026
Developer: DeepSeek (Hangzhou)
License: MIT
Hugging Face: `deepseek-ai/DeepSeek-V4-Pro` and `deepseek-ai/DeepSeek-V4-Flash`

Architecture & Key Specs

DeepSeek V4 shipped as a two-tier lineup — a first among frontier open-weight releases:

• V4-Pro: 1.6 trillion total parameters, ~49 billion activated per token, 1 million token context window

• V4-Flash: 284 billion total parameters, ~13 billion activated per token, same architecture, single 80GB GPU compatible when quantized

The architecture introduces DeepSeek Sparse Attention (DSA) — selecting only top-k compressed tokens for full attention computation while keeping a 128-token sliding window for local context. The result: V4-Pro cuts single-token inference FLOPs to 27% of V3.2, and KV cache occupancy to just 10% of V3.2 at the 1M-token setting.

Benchmark Performance

DeepSeek V4-Pro scored 52 on the Artificial Analysis Intelligence Index — a 10-point jump from V3.2’s score of 42. It became the #2 open-weight reasoning model, trailing only Kimi K2.6 (54) on that index.

On the Artificial Analysis Coding Index, V4 sits “very close to the previous generation of frontier models,” per independent analysis. The model achieves competitive results on SWE-Bench while delivering dramatically lower inference costs.

Real-World Inference Cost Advantage

According to DeepSeek’s own documentation, running V4-Pro at 1M-token context now costs roughly 27% of the FLOPs required for equivalent V3.2 inference. For developers running long-document workflows, code repository analysis, or agentic pipelines, this translates to direct dollar savings.

> Key stat: V4-Pro is the first open-weight model to combine 1M-token context, sub-50B active parameters, and MIT licensing — a combination that makes local deployment on enterprise hardware genuinely practical.

Availability

• API: Available via `api.deepseek.com` (preview launched April 24)

• Open weights: Hugging Face (`deepseek-ai/DeepSeek-V4-Pro`, `deepseek-ai/DeepSeek-V4-Flash`)

• GGUF builds: Available via Unsloth for consumer GPU deployment

—

4. Kimi K2.6: The Coding & Agent Swarm Champion

Release Date: April 20, 2026
Developer: Moonshot AI (Beijing)
License: Modified MIT License
Hugging Face: `moonshotai/Kimi-K2.6`

Architecture & Key Specs

Kimi K2.6 is a 1-trillion-parameter Mixture-of-Experts model with 32 billion active parameters per token, released open-weight under a Modified MIT License. Four variants serve different use cases:

1. Instant — Fast responses for casual chat
2. Thinking — Deep reasoning for complex problems
3. Agent — Autonomous research and document tasks
4. Agent Swarm — Large-scale parallel work with up to 300 sub-agents and 4,000 coordinated steps

The sparse MoE design uses 384 total experts, routing 8 active experts plus 1 shared expert per token — the same core architecture as K2 and K2.5, but with significantly enhanced training.

Benchmark Performance

Kimi K2.6 set new records on coding benchmarks:

| Benchmark | K2.6 | GPT-5.4 (xhigh) | Claude Opus 4.6 | Gemini 3.1 Pro |
|———–|——|—————–|—————–|—————-|
| SWE-Bench Pro | 58.6% | 57.7% | 53.4% | 54.2% |
| HLE-Full (w/ tools) | 54.0 | 52.1 | 53.0 | 51.4 |
| DeepSearchQA (accuracy) | 83.0% | 63.7% | 80.6% | 60.2% |

On SWE-Bench Pro coding benchmarks, K2.6 ties GPT-5.5 at 58.6% — a remarkable result for an open-weight model. On the Artificial Analysis Intelligence Index, K2.6 leads all open-weight models at 54 points.

The Agent Swarm: Defining Capability

The most distinctive feature of K2.6 is Agent Swarm (branded “Claw Groups” in partner ecosystems). This allows the model to scale to 300 parallel sub-agents coordinating through 4,000 sequential steps. In real-world terms, this means:

• An entire software sprint can be planned, coded, tested, and reviewed autonomously

• Multi-repository refactoring at scale becomes feasible

• Research agents can canvas thousands of sources simultaneously

Moonshot claims code generation accuracy increased by 12%, long-context stability improved by 18%, and tool invocation success reached 96.60% compared to K2.5 in internal CodeBuddy evaluations.

> Key stat: Kimi K2.6 achieved an 83.0% DeepSearchQA accuracy — vs. GPT-5.4’s 63.7% and Gemini 3.1 Pro’s 60.2%. For open-source research tasks, this gap is enormous.

Availability

• API: `platform.moonshot.ai` with OpenAI-compatible endpoint

• Open weights: Hugging Face (`moonshotai/Kimi-K2.6`)

• Third-party API: Available via DeepInfra, OpenRouter, and other providers

—

5. GLM-5.1: Agentic Engineering Powerhouse

Release Date: Early May 2026 (open-source weights)
Developer: Z.ai (Zhipu AI, Beijing) — China’s first public AI company
License: MIT
Hugging Face: Weights on Hugging Face via Z.ai

Architecture & Key Specs

GLM-5.1 is a 744-billion-parameter MoE model with approximately 40 billion active parameters per token, designed specifically for agentic engineering — Z.ai’s term for autonomous, long-horizon task completion in complex software systems.

The model builds on GLM-5 (released February 2026), refining the architecture for extended autonomous execution. GLM-5.1 is capable of working on a single coding task for up to eight hours autonomously — a spec that directly addresses enterprise agentic workflows.

Benchmark Performance

GLM-5.1 achieved SOTA (State of the Art) on SWE-Bench Pro with a score of 58.4%, surpassing GPT-5.4 (57.7%), Claude Opus 4.6 (53.4%), and Gemini 3.1 Pro (54.2%):

| Model | SWE-Bench Pro Score |
|——-|———————|
| GLM-5.1 | 58.4% (SOTA) |
| Kimi K2.6 | 58.6% |
| GPT-5.4 | 57.7% |
| Claude Opus 4.6 | 53.4% |
| Gemini 3.1 Pro | 54.2% |

Z.ai’s own documentation claims GLM-5.1 is “aligned with Claude Opus 4.6.” Independent evaluations indicate GLM-5.1 achieves approximately 94.6% of Claude Opus 4.6’s coding score, with remaining gaps in reasoning and creative tasks. But on SWE-Bench Pro — the gold-standard coding benchmark — GLM-5.1 leads among all models tested.

Agentic Engineering Focus

GLM-5.1’s defining capability is its design for agentic engineering — the model is specifically trained to:

• Execute multi-step software engineering tasks autonomously

• Maintain context across 8-hour coding sessions

• Interface reliably with external tools and APIs

• Operate within complex engineering system environments

> Key stat: GLM-5.1 scored 58.4% on SWE-Bench Pro — the highest score ever recorded at time of release, beating every closed-source frontier model including GPT-5.4 and Claude Opus 4.6.

Availability

• API: Available via `api.z.ai` (Z.ai Open Platform)

• Open weights: Open-source release on Hugging Face (May 2026)

• Note: Zhipu AI has been on the U.S. Entity List since January 2025, restricting access to H100/H200 GPUs — making this model’s performance under those constraints particularly notable

—

6. MiniMax M2.7: Self-Evolving Open Model

Release Date: Open weights April 12, 2026 (announced March 18, 2026)
Developer: MiniMax (Shanghai)
License: Custom open-weight license (not fully Apache)
Hugging Face: `MiniMaxAI/MiniMax-M2.7`

Architecture & Key Specs

MiniMax M2.7 is a 230-billion-parameter sparse MoE model with only 10 billion active parameters per token, making it the most parameter-efficient model in this comparison. It’s the first MiniMax model to “actively participate in its own development cycle” — a self-evolving agent model.

The self-evolution cycle means: during development, M2.7 updated its own memory, built dozens of complex skills for RL experiments, and improved its own learning process based on experiment results. This is a meaningful shift in how large language models are built and iterated.

Core capability areas:

• Professional software engineering

• Professional office work

• Agent Teams — native multi-agent collaboration

Benchmark Performance

| Benchmark | MiniMax M2.7 Score |
|———–|——————–|
| SWE-Bench Pro | 56.22% |
| Terminal Bench 2.0 | 57.0% |
| VIBE-Pro (repo-level code gen) | 55.6% |

VIBE-Pro at 55.6% puts M2.7 “nearly on par with Opus 4.6” for repository-level code generation tasks covering Web, Android, iOS, and simulation tasks.

The Self-Evolution Innovation

MiniMax’s approach to M2.7 was fundamentally different from traditional model development:

1. The model built its own memory mechanisms during training
2. It constructed complex skills autonomously for reinforcement learning experiments
3. It improved its own learning process based on results

This self-evolution approach resulted in an agent model that adapts to new tasks more fluidly than traditionally trained models. NVIDIA integrated high-performance kernels for M2.7 into vLLM (2.5x throughput improvement) and SGLang (2.7x throughput improvement) in collaboration with the open-source community.

> Key stat: M2.7 scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2.0 — and it’s built around a 10B-active-parameter architecture, making it the most deployable of the four models for single-GPU scenarios.

Availability

• Open weights: Hugging Face (`MiniMaxAI/MiniMax-M2.7`)

• Inference: Recommended via SGLang or vLLM

• Note: The license is not fully Apache — check Hugging Face repository for specific terms

—

7. Head-to-Head Benchmark Comparison

Here’s how all four models stack up across the key benchmarks that matter for AI practitioners:

| Model | SWE-Bench Pro | Intelligence Index | Architecture | Active Params | Context Window | License |
|——-|:————-:|:—————–:|:————:|:————-:|:————–:|———|
| Kimi K2.6 | 58.6% | 54 | 1T MoE | 32B | 128K+ | Modified MIT |
| GLM-5.1 | 58.4% | ~52 | 744B MoE | 40B | 128K+ | MIT |
| DeepSeek V4-Pro | ~52%* | 52 | 1.6T MoE | 49B | 1M | MIT |
| MiniMax M2.7 | 56.22% | ~50 | 230B MoE | 10B | 128K+ | Custom |

*DeepSeek V4 benchmark on SWE-Bench Pro estimated from available data — official scores vary by evaluation methodology.

Key Takeaways

• Best raw coding performance: Kimi K2.6 and GLM-5.1 are essentially tied at the top (58.6% vs 58.4%)

• Best long-context capability: DeepSeek V4-Pro wins with 1M token context (vs 128K for others)

• Most deployable (smallest active params): MiniMax M2.7 at 10B active parameters

• Best inference efficiency: DeepSeek V4-Pro at 27% FLOPs vs V3.2

• Most agentic (multi-agent): Kimi K2.6 with Agent Swarm (300 sub-agents)

• Self-evolution: MiniMax M2.7 is the only model that participated in its own training

—

8. Inference Cost: Why Western Models Can’t Compete

This is where Chinese open-weight models pull decisively ahead of their Western counterparts.

Comparative Inference Cost Analysis

A developer running a coding agent pipeline with 10,000 requests/month:

| Model | Active Params | Inference Cost (approx.) | Monthly Cost @ 10K req |
|——-|:————-:|————————|———————-|
| GPT-5.4 (xhigh) | ~200B+ | ~$15/1M tokens | ~$150–300 |
| Claude Opus 4.6 | ~200B+ | ~$15/1M tokens | ~$150–300 |
| Kimi K2.6 | 32B | ~$0.5–1/1M tokens | ~$5–10 |
| GLM-5.1 | 40B | ~$0.5–1/1M tokens | ~$5–10 |
| DeepSeek V4-Pro | 49B | ~$0.3–0.7/1M tokens | ~$3–7 |
| MiniMax M2.7 | 10B | ~$0.2–0.5/1M tokens | ~$2–5 |

That’s a 20–50x cost advantage for developers choosing Chinese open-weight models over Western closed APIs.

The lower inference cost stems from three structural advantages:

1. Sparse MoE architecture — only a fraction of parameters activate per token
2. Smaller active parameter counts — MiniMax M2.7’s 10B active params vs 200B+ for closed models
3. Competitive domestic pricing — Chinese API providers price aggressively to gain market share

For a developer building an AI coding assistant used by 1,000 monthly active users, switching from GPT-5.4 API to Kimi K2.6 or GLM-5.1 could save $10,000–$50,000 per month.

—

9. Who Should Use Which Model?

Choose DeepSeek V4-Pro if:

• You need 1M token context for long-document analysis, entire codebases, or legal document review

• Inference cost efficiency is your #1 priority

• You want a two-tier model (Pro for capability, Flash for speed/cost)

• You’re running on-premise with consumer or enterprise GPU hardware

• You want MIT licensing for unrestricted commercial use

Choose Kimi K2.6 if:

• You need state-of-the-art coding performance (ties GPT-5.5 on SWE-Bench)

• You’re building multi-agent systems requiring Agent Swarm or Claw Groups

• You need high accuracy on research and deep-search tasks (83.0% DeepSearchQA)

• You want OpenAI-compatible API for easy migration from GPT-4

• You’re building autonomous coding agents or developer tooling

Choose GLM-5.1 if:

• You prioritize SWE-Bench Pro leadership (58.4%, the SOTA score)

• You’re building agentic engineering systems that run for hours autonomously

• You want MIT licensing with strong open-source ecosystem support

• You’re comparing against Claude Opus 4.6 and want comparable performance at 1/20th the cost

• You need a model specifically designed for complex software engineering workflows

Choose MiniMax M2.7 if:

• You want the smallest active parameter footprint (10B active) for maximum deployability

• You’re building multi-agent teams with native collaboration primitives

• You want a model that can adapt and evolve its own capabilities

• You need strong performance on both software engineering AND office work tasks

• You’re running on limited GPU resources (single 80GB GPU or lower)

—

10. The Bottom Line

The April–May 2026 open-weight blitz from Chinese AI labs fundamentally changed the economics of AI development. Four models released in 12 days — DeepSeek V4, Kimi K2.6, GLM-5.1, and MiniMax M2.7 — delivered frontier-class performance on coding, reasoning, and agentic workflows at 20–50x lower inference cost than Western closed models.

The implications are significant:

1. Open-source AI is now competitive at the frontier — these models match or beat GPT-5.4 and Claude Opus 4.6 on key benchmarks
2. Agentic AI is production-ready — Agent Swarm, self-evolution, and 8-hour autonomous coding sessions are no longer research demos
3. Inference economics have shifted permanently — a $50,000/month API bill can become a $2,000/month self-hosted deployment
4. Chinese AI labs are now leading — not just matching — on open-weight model capability

For developers, startups, and enterprises, now is the time to evaluate these models for your specific use case. The cost savings alone justify a migration evaluation — and the performance data suggests you won’t be sacrificing quality.

Start exploring today: All four models are available on Hugging Face and through various API providers. Your next production AI agent might cost 20x less than it does today.

—

*Ready to dive deeper? Explore related guides on [AI coding agents that actually make money](https://yyyl.me), [the best open-weight models for side hustles](https://yyyl.me), and [how to build your first AI agent pipeline](https://yyyl.me).*