1. Overview

This is the third article in the "ALL IN AI" column, and I want to discuss the latest progress of domestic models. To be honest, I originally didn't plan to write this kind of news roundup article. Simply tracking model updates is actually quite limited in helping improve personal skills. However, domestic models have developed very rapidly in the last two months. Some friends might have missed many important changes because they were busy with work or don't usually follow this field. The main reason is that in the next article, I plan to introduce how to connect Claude Code and Codex to domestic models. Before getting into the specific operations, it's necessary to first sort out what has happened recently with DeepSeek, Kimi, and GLM, which also serves as a foundation for the upcoming practical tutorial.

Recently, the update speed of domestic large models has been very fast: DeepSeek V4 Flash 0731 and DeepSeek V4 Pro 0813 were officially released, Kimi K3 officially debuted with 2.8T parameters, native vision, and 1M context, and the GLM Coding Plan, which was once "impossible to get," has finally fully opened for subscription. Looking at these changes together points to an increasingly clear trend: The gap between domestic models and top overseas models is rapidly narrowing, and they have already formed their own competitiveness in terms of price, context length, and accessibility.

This article focuses on three questions:

• What noteworthy changes have occurred recently with DeepSeek V4, Kimi K3, and GLM?
• To what extent has the gap between domestic models and top overseas models narrowed?
• Why can't you judge a model's "intelligence" based on just one or two conversations?

It must be emphasized here that this article does not conduct any subjective model "beatdown list" evaluations. When it comes to capability comparisons, it mainly cites official data from manufacturers' websites and third-party evaluations like Artificial Analysis; price information and comparisons are based on official materials available as of the writing date of this article, August 14, 2026.

2. What is the actual level of domestic models now?

If you only look at social media, it's easy to swing between two extremes: one voice believes domestic models have already comprehensively surpassed overseas models; another voice believes domestic models can still only handle simple tasks. I think neither of these views is objective enough. Those who think domestic models are still at a beginner level likely have strong subjective bias, believing that domestic products are inherently inferior to foreign ones (PS: Many people say Huawei phones are just not as good as iPhones~). That's not the case; only what truly suits you is the best. But conversely, saying domestic models have already comprehensively surpassed overseas models is also somewhat overconfident, a bit like some domestic media that like to exaggerate facts and blindly boast. Anthropic and OpenAI are still the leaders in the AI field, with deep accumulation in model R&D, computing power investment, product ecosystem, and engineering capabilities. Just in terms of their computing power and GPU configurations, domestic players certainly can't compare. So my judgment is simple: In terms of overall performance, domestic models have not yet comprehensively surpassed top overseas models, but the gap between the two sides is indeed rapidly narrowing.

Model capability is not a single number. Reasoning, knowledge, code, vision, tool use, long context, generation speed, and stability can all yield different rankings. Even using the same model, switching an Agent, a set of tools, or a Prompt can lead to completely different results.

But one point is becoming increasingly hard to deny: domestic models are no longer just "cheap, elementary-level substitutes."

• DeepSeek V4 Pro and V4 Flash both offer 1M context, thinking mode, and tool-calling capabilities;
• Kimi K3 scales up to 2.8T parameters and simultaneously covers long tasks, knowledge work, and native vision;
• GLM-5.2 has entered the first tier of open-weight models in third-party comprehensive evaluations;
• Multiple domestic model manufacturers have begun actively adapting to Agent clients like Claude Code, Codex, and OpenCode;
• Model prices, cache prices, and RMB payment methods significantly lower the barrier for high-frequency use compared to overseas models.

The chart below first establishes an overall impression of the three product lines:

It's important to note here that parameter scale, context length, and price cannot be directly equated to "intelligence." They only indicate the model's product positioning and usable boundaries; true capability still needs to be judged in combination with specific tasks and testing conditions.

3. DeepSeek V4 Pro and V4 Flash

3.1 What scenarios are the two versions aimed at?

DeepSeek V4 is divided into Pro and Flash versions. The current official API versions are DeepSeek-V4-Pro-0813 and DeepSeek-V4-Flash-0731:

Model	Total Params / Active Params	Main Positioning	Context	Max Output
DeepSeek V4 Pro	1.6T / 49B	Complex reasoning, long tasks, and higher delivery quality	1M	384K
DeepSeek V4 Flash	284B / 13B	Faster response, higher concurrency, and lower cost	1M	384K

Both support:

• Non-thinking mode and thinking mode;
• Reasoning intensity levels like high, max;
• Tool Calls;
• JSON output;
• Conversation prefix continuation;
• OpenAI and Anthropic compatible interfaces.

DeepSeek's official positioning is intuitive: V4 Pro handles more difficult, longer tasks requiring stable reasoning, while V4 Flash pushes speed and cost lower while retaining strong Agent capabilities.

This doesn't mean all complex tasks must use Pro, nor that Flash can only complete simple tasks. Model selection is better dynamically decided based on task risk, execution length, latency requirements, and budget.

3.2 Why is V4 Flash attracting attention?

V4 Flash's attention comes partly from its Agent capabilities and partly from its price.

According to evaluation data disclosed at DeepSeek's launch, V4 Flash achieved very competitive results in tests like Terminal-Bench 2.1, DeepSWE, Toolathlon-Verified, and DSBench-FullStack, surpassing the previous V4 Pro preview version in some projects. However, such results first represent the performance of a specific model version under specific Harnesses, reasoning tiers, and parameter configurations, and cannot be directly extrapolated to mean "Flash is better than the Pro preview in all real projects."

What's truly noteworthy is: Low-priced models no longer necessarily equate to only handling low-difficulty tasks.

3.3 DeepSeek V4 Pricing

As of August 14, 2026, DeepSeek's current RMB prices are as follows:

Per Million Tokens	V4 Flash	V4 Pro
Cache Hit Input	0.02 RMB	0.025 RMB
Cache Miss Input	1 RMB	3 RMB
Output	2 RMB	6 RMB
Account Concurrency Limit	2500	500

In this price structure, the most noteworthy aspect is the cache hit input. Agents repeatedly carry system instructions, historical messages, and project context when handling long tasks. If the cache can be consistently hit, input costs will drop significantly.

Of course, whether the cache is hit is not determined by "the context looks similar." Session switching, model switching, reasoning tier changes, or context rewriting can all invalidate the cache.

Special attention: DeepSeek has announced that a new peak/off-peak pricing will be activated on August 17, 2026, at 00:00. At that time, Beijing time 9:00–12:00 and 14:00–18:00 will be peak hours, and the rest will be off-peak hours:

Model	Period	Cache Hit Input	Cache Miss Input	Output
V4 Flash	Off-peak	0.05 RMB	1.5 RMB	4.5 RMB
V4 Flash	Peak	0.10 RMB	3 RMB	9 RMB
V4 Pro	Off-peak	0.15 RMB	4.5 RMB	13.5 RMB
V4 Pro	Peak	0.30 RMB	9 RMB	27 RMB

In other words, "peak/off-peak pricing" does not simply mean a discount on the current price. After August 17, off-peak prices are half of peak prices, but both periods have changed compared to the prices in effect on August 14. Users planning long-term API usage should refer to the official pricing page on the day of the call.

4. Kimi K3: 2.8T Parameters, Native Vision, and 1M Context

In July 2026, Kimi officially released Kimi K3. It is positioned as a flagship model for long-range tasks, knowledge work, and complex reasoning.

K3's most prominent numbers are:

• 2.8T total parameters;
• 1M Token context;
• Native visual understanding;
• Supports low, high, max reasoning intensities;
• Geared towards Agent workflows like Kimi Code, Claude Code, Codex;
• Model weights have been opened.

Compared to text-only models, native vision allows K3 to directly understand screenshots, charts, interfaces, and documents containing images. For long tasks requiring simultaneous processing of text, pages, and visual information, this is a very practical capability.

4.1 Long context isn't just about "fitting more text"

1M context means the model can receive more requirements, documents, code, terminal output, and history at once. But "fitting it in" doesn't equal "using it well."

The real difficulties of long tasks include:

• Remembering the initial goal after many rounds of operation;
• Finding the currently relevant information from a vast context;
• Not letting error logs or repeated output crowd out effective context;
• Continuing the task direction after compressing the session;
• Being able to revise previous plans based on test or tool results.

Kimi introduced Kimi Delta Attention and Attention Residuals in K3, aiming to improve information utilization and computational efficiency in ultra-long sequences. Ordinary users don't need to remember these architectural terms; they just need to know: K3's product focus is not single-round answers, but enabling the model to work continuously on longer, more complex tasks.

4.2 K3 and K3-256k

Kimi Code currently offers two related model IDs: k3 and k3-256k:

Model ID	Context	Applicable Situations
k3	1M	Ultra-long materials, very large projects, and long-duration tasks
k3-256k	256K	Daily Q&A, routine modifications, and tasks that save quota

According to the official description, K3-256k provides the same model capability within the 256K range but consumes less package quota. Most daily tasks may not necessarily require using 1M directly; having enough context is more important than the maximum number.

4.3 K3 Pricing

The official K3 API prices announced by Kimi are:

Per Million Tokens	Kimi K3
Cache Hit Input	0.30 USD
Cache Miss Input	3.00 USD
Output	15.00 USD

K3's unit price is significantly higher than DeepSeek V4, but it offers native vision and a higher flagship positioning. Whether it's worth using depends on whether the task truly needs these capabilities. K3's pricing is not lower than top models like Claude~

5. GLM Coding Plan Finally Fully Open for Subscription

The most prominent problem with the GLM Coding Plan previously wasn't that it was expensive, but that it was unavailable. Due to computing power and supply constraints, the plan was in a limited state for a long time, and new users might not be able to subscribe even if they were willing to pay. Now this problem is finally solved: the GLM Coding Plan is fully open, and both new and existing users can subscribe directly.

But the accompanying change is also very obvious—the price has increased.

5.1 Old vs. New Price Comparison

Plan	Old Monthly Fee	New Continuous Monthly	Increase
Lite	49 RMB	118 RMB	~141%
Pro	149 RMB	538 RMB	~261%
Max	469 RMB	1078 RMB	~130%

Looking at the monthly fees alone, none of the three tiers are cheap, and the change for Pro is particularly noticeable.

The new plans also adjust the usage calculation method. The official site counts available usage based on 5 hours and a weekly resource pool, and different models are deducted at different coefficients. High-tier models like GLM-5.2 and GLM-5-Turbo have different consumption rates during peak and off-peak hours, with peak hours defined as 14:00–18:00 daily (UTC+8).

The GLM Coding Plan currently supports GLM-5.2, GLM-5-Turbo, and GLM-4.7, and provides exclusive MCP capabilities like visual understanding, web search, webpage reading, and open-source repository access.

So the significance of this full opening is very clear: The problem of "wanting to buy but unable to" is solved, but it is no longer a plan you can subscribe to with your eyes closed just by looking at the price.

Light users need to estimate their own frequency first; heavy users of Agents like Claude Code and OpenCode can judge whether it's cost-effective based on the weekly quota, model deduction coefficients, and peak hour rules.

6. How big is the gap between domestic models and top overseas models?

To discuss this question, it's best to first acknowledge a fact: there is no single score that can summarize all of a model's capabilities.

To make the data as comparable as possible, only the Artificial Analysis Intelligence Index v4.1.1 is used below. This index synthesizes nine evaluations including GDPval-AA v2, Terminal-Bench 2.1, SciCode, Humanity's Last Exam, and GPQA Diamond.

The chart uses the corresponding high reasoning tiers for each model, with the specific tiers marked directly in the model names. According to the data available as of August 14, 2026:

• Claude Fable 5 (max, including Opus 4.8 fallback) scored 60;
• GPT-5.6 Sol (max) scored 59;
• Kimi K3 (max) scored 57;
• DeepSeek V4 Pro 0813 (max) scored 53;
• DeepSeek V4 Flash 0731 (max) scored 52;
• GLM-5.2 (max) scored 51;

This set of data at least shows that domestic models can now enter the same high-level capability chart for comparison. Kimi K3 is only 3 points behind the highest score, and the newly updated DeepSeek V4 Pro 0813 and V4 Flash 0731 have also entered the 50+ point range.

But this chart cannot explain the following:

• Kimi K3 is overwhelmingly strong in all scenario tasks;
• GLM-5.2 is definitely more suitable for your work than DeepSeek V4 Pro;
• Models with a few points lower score cannot complete complex tasks;
• Model price, speed, visual ability, and context management are unimportant.

Furthermore, model scores on manufacturer release pages may use different Agent Harnesses. Kimi K3 might be paired with Kimi Code, Claude models might be paired with Claude Code, and GPT models might be paired with Codex. The Harness determines which tools the model can use, how context is managed, and how execution results are returned, so it is inherently part of the final score.

Let's look at another comparison chart:

This is a vivid saying that has become popular recently—"the large model execution line."

If we treat an Artificial Analysis Intelligence Index score of 50 as the capability threshold for a strong model, and a single-task cost of 0.03 USD as a price reference line, then DeepSeek V4 Flash almost falls right at the intersection of these two lines. Its composite score may not be the highest, but among models reaching the same capability range, its cost is outstandingly low.

A more accurate conclusion should be: There are still differences between domestic models and top overseas models, but the gap is not as large as many people imagine; in some testing and usage dimensions, domestic models already have leading or near-leading performance.

7. Why can't you judge a model's quality based solely on personal feeling?

"This model has gotten dumber."

"That model can't write code."

"I tried it once, completely unusable."

These experiences are all real, but they are usually insufficient to judge a model's overall capability.

The result of the same task is at least influenced by the following factors:

1. Is the task clearly stated: Are the goals, boundaries, inputs, and acceptance criteria complete;
2. Is the context sufficient: Has the model seen the truly necessary files and information;
3. Are the tools suitable: Can the Agent read files, execute commands, access the internet, or view images;
4. Does the reasoning budget match: Turning on the highest tier for simple tasks wastes cost, while too low a budget for complex tasks may lead to hasty conclusions;
5. Is the execution process verifiable: Are there tests, data, or clear standards to tell the model where it went wrong;
6. Is the model itself suitable for the task: Different models indeed have different capability focuses.

A poorly written Prompt can certainly make a model appear not smart enough. For example, just saying "help me optimize it" doesn't tell the model whether to optimize speed, cost, structure, or expression, nor what result counts as completion.

But not all failures can be blamed on the user not knowing how to write Prompts. It's also possible that the model's capability is insufficient, leading to misunderstanding, forgetting goals, incorrectly calling tools, hallucinating, or simply lacking the visual and knowledge abilities required for the task.

A more reliable judgment method is:

First look at public evaluations with clear sources, then have different models complete the same real task, and finally choose based on a combination of quality, speed, stability, and cost.

A model's quality is not determined by someone's single sentence, but reproducible results in real tasks are still more important than marketing slogans.

8. Summary

The recent changes in DeepSeek V4, Kimi K3, and the GLM Coding Plan indicate that domestic models are entering a new phase.

From public data, domestic models and top overseas models still have their respective strengths, but the gap has significantly narrowed:

• Kimi K3 has entered the global first tier in some third-party composite indicators;
• GLM-5.2 has become a very competitive open-weight model;
• DeepSeek V4 Pro offers complex task capability with very low call costs;
• DeepSeek V4 Flash further lowers the price threshold for high-frequency tasks.

Whether a model is good cannot be judged solely by manufacturer launch events, nor by a single person's one-time experience. Prompts, tools, context, and task verification all affect the results, and the model's own capability boundaries also objectively exist.

For ordinary users, the more valuable question is not "who is the single strongest model," but:

Within your own usage scenario and budget, which model can efficiently and stably complete real tasks?