China's LLM market is entering a dense release window in July 2026, with three major models converging on the same month: DeepSeek V4 (full release), Kimi K3, and Zhipu GLM-5.3. What unites them is a shared push across coding capability, agent engineering, and long-context understanding — marking the transition from "chasing benchmarks" to "getting real work done."
This article breaks down the known highlights of each model and why they matter.
---
1. DeepSeek V4: The Open-Source Efficiency Benchmark
MoE Architecture: 49B Active Parameters Delivering SOTA Performance
According to DeepSeek's official announcement, the V4 preview was released on April 24, with the full version targeted for mid-July. It uses a dual-tier Mixture-of-Experts (MoE) architecture:
| Variant | Total Params | Active Params | Pretrain Tokens | Context |
|---|---|---|---|---|
| V4-Pro | 1.6T | 49B | 33T | 1M |
| V4-Flash | 284B | 13B | 32T | 1M |
V4-Pro-Max took first place in 5 out of 7 major benchmarks. Codeforces score: 3206. SimpleQA: 57.9. SWE Verified: 80.6 — on par with top closed-source models in software engineering agent capability.
What's remarkable is that this was achieved with only 49B active parameters — the MoE architecture's efficiency advantage is on full display.
DSpark Inference Acceleration Framework
The DSpark paper published on June 28 (DeepSeek official blog) reveals V4's secret weapon: an inference acceleration system already deployed in V4-Pro and V4-Flash production environments.
Key metrics:
- Single-user generation speed improved by 60%–85%
- System throughput in high-concurrency scenarios increased severalfold
DSpark's genius lies in the co-design of systems engineering and models: speculative decoding, parallel draft models, confidence scheduling, hardware-aware scheduling — no single technique is novel, but assembling them into a stable production system is where the real engineering capability shows.
Full Release vs Preview: A Leap in Coding Capability
The preview already showcased V4's parameters and benchmarks, but what has the community excited is the potential capability improvement in the full release. Recent users have reported receiving gray-box API test invitations for V4 full release, and the shared experience points to coding capability as the most significant improvement — better comprehension of complex multi-file projects, improved consistency in long-context code generation, and enhanced stability in multi-step tool-calling under agent mode.
DeepSeek's own team posted user feedback surveys on social platforms to collect pain points and improvement suggestions from V4 real-world usage, signaling that the full release targets the preview's identified weaknesses.
Why it's worth watching: If you're already using V4 preview, the full release's performance in coding-agent scenarios deserves close attention. The combination of 49B active parameters + DSpark acceleration, plus the full release's capability tuning, could make this the best value-for-money coding model available.
References: DeepSeek API Docs ｜ DeepSeek on HuggingFace
---
2. Kimi K3: The Linear Attention Efficiency Revolution
Trillion Parameters + Linear Attention
Yang Zhilin confirmed at the AGI-next forum that Kimi K3 will be released in July, and further revealed technical directions during a Reddit AMA. Based on these disclosures, the core profile is:
- Parameter scale: Expanded from K2's 1 trillion parameters
- Core breakthrough: Improvements in linear attention mechanisms, with inference speed potentially improving 6–10×
- Context window: Expected to maintain 1M tokens
Linear attention is the current bottleneck in LLM inference efficiency. Standard attention has O(n²) complexity — the longer the sequence, the more exponentially the compute cost rises. If K3 can reduce this to near-linear, it means 1M context becomes not just "feasible but expensive" but "feasible and affordable."
Agent-First Product Philosophy
Kimi has been shipping agent products throughout 2026: Researcher, OK Computer, PPT generation, and Kimi Code. K3's training objective is clear — not a general-purpose chatbot, but an agent foundation model.
The K2 Thinking release already validated this approach: it trended globally on Twitter, and overseas API revenue quadrupled in a single quarter. If K3 can further improve long-horizon task success rates on this foundation, it would be a genuinely agent-native model.
Commercial Validation
Kimi's latest funding round valued the company at $31.5 billion pre-money, with ARR tripling in three months and API revenue accounting for over 70% of total. The market has voted with real money.
Why it's worth watching: If the linear attention efficiency gains materialize, K3 could be the first trillion-parameter model to achieve a qualitative leap in inference cost — with implications for the entire industry's pricing structure.
References: Kimi Official ｜ Moonshot Platform
---
3. GLM-5.3: Zhipu's Agent Engineering Path
Reading the Trend from GLM-5.2
To understand GLM-5.3's direction, start with GLM-5.2, released in June:
- 1M context window
- High/Max dual thinking modes (effort-level control, letting users choose reasoning depth)
- MIT-licensed open source
GLM's core positioning has been consistent: Coding + Agentic Engineering. Version 5.2 isn't about stacking benchmark scores — it's about deep optimization for complex engineering agent scenarios: ultra-long codebase comprehension, multi-step task orchestration, and tool-calling reliability.
In real-world developer community feedback, GLM-5.2's coding capability has been rated as "on par with top closed-source models" — exceptionally high praise for an MIT-licensed open model.
GLM-5.3: What to Expect
Zhipu CEO Tang Jie initiated a next-generation model feedback survey on X (Twitter), with discussions spreading to platforms like Xiaohongshu. Based on information surfaced in these threads, two directions deserve attention:
- Multimodal visual capability: GLM-5.2 has proven its text and code prowess. Adding visual understanding is the natural next step — agents need to "see": read screenshots, understand UI layouts, parse chart data.
- Deeper agent orchestration: When context windows are no longer the bottleneck, competition shifts to: can the model maintain reasoning quality across 1M tokens? Can it sustain task consistency across dozens of tool-call rounds?
Zhipu's advantage: they started early on agent engineering, and the accumulated know-how is more valuable long-term than benchmark numbers.
Why it's worth watching: The GLM series has always been the "quiet but reliable" presence in the open-source ecosystem. If 5.3 can break through on multimodality, combined with existing agent engineering capabilities, it would form a very complete product.
References: Zhipu AI Open Platform ｜ Zhipu AI Official
---
Common Trend: From "Chasing Scores" to "Getting Things Done"
Looking at all three together, a clear consensus emerges:
The competitive focus has shifted from benchmarks to agent capability.
- DeepSeek V4 leverages MoE efficiency + DSpark acceleration to drive down inference cost
- Kimi K3 uses linear attention to attempt a speed revolution at trillion-parameter scale
- GLM-5.3 likely doubles down on agent engineering + multimodality
All three converge on the same direction: making AI that actually works — writing code, running tasks, processing ultra-long documents, making autonomous decisions.
For developers, this means real capability dividends in the second half of 2026.
---
Access All Three Models Through SpanAgent
If you want to integrate these models into your projects, SpanAgent provides a unified API aggregation service:
- DeepSeek V4 (V4-Pro / V4-Flash) — pay-per-use, with peak/off-peak pricing
- Kimi K3 (will be integrated immediately upon release) — natively compatible with OpenAI API format
- GLM series (currently supports GLM-5.2, 5.3 integrated upon release) — full API support for MIT-licensed open models
SpanAgent advantages:
- Unified API format — integrate once, call all models
- Pay-as-you-go — no prepaid plans required
- Smart routing — automatically selects the fastest channel
- Full request logging and usage analytics
Start now: https://spanagent.xyz
---
Information current as of early July 2026. Model parameters and release dates are subject to official announcements from each vendor.