DeepSeek V4 and V4 Flash vs. Kimi K3: How the New Wave of Open Models Stacks Up

If you have been anywhere near tech news in the last few months, you have probably seen a flood of headlines about new AI models with names like DeepSeek V4, V4 Flash, and Kimi K3. It can be hard to tell what is actually new, what matters, and why anyone should care. This post is a plain-language tour of the latest crop of open models, what makes them different, and where each one shines. No PhD required.

First, what does "open" even mean here?

When people say a model is "open weight," they mean the company has published the actual trained model files so anyone can download them, run them on their own hardware, and often build products on top of them. This is different from a closed model like the ones behind many popular chat apps, where you can only reach the model through a paid connection and never see what is inside. Open weight does not always mean fully open source in the strict software sense, and licenses vary, but the practical upshot is the same: more control, more privacy options, and usually far lower costs. That shift is a big part of why these releases are getting so much attention.

DeepSeek V4 and V4 Flash: two sizes, one family

DeepSeek launched its V4 generation in late April 2026 as a two-model family, both released under the permissive MIT license and both offering an unusually large one-million-token context window. Think of the context window as the model's short-term memory: a million tokens is roughly enough to hold several long books at once, which is handy for working across big codebases or long documents.

The bigger sibling is DeepSeek V4 Pro, a Mixture-of-Experts model with about 1.6 trillion total parameters but only around 49 billion active at any given moment. The lighter sibling is V4 Flash, which carries 284 billion total parameters and about 13 billion active. If that jargon is unfamiliar, here is the short version: Mixture-of-Experts means the model is made of many specialist sub-networks, and only a few of them switch on for each request. That is why Flash can be fast and cheap to run while still being capable.

The story got more interesting on July 31, 2026, when DeepSeek shipped an updated build of V4 Flash, often referred to by its "0731" tag. According to DeepSeek's own model card, this version kept the same size and architecture as the earlier preview but reran the final training stage to sharpen its skills at coding and agentic tasks, meaning tasks where the model uses tools and works through multi-step problems on its own. The gains were large. On the company's published benchmarks, the new Flash jumped from 61.8 to 82.7 on Terminal-Bench 2.1, a test of command-line problem solving, and it climbed steeply on several coding evaluations compared with the earlier preview.

The other headline is price. DeepSeek's API pricing for Flash sits around 14 cents per million input tokens and 28 cents per million output tokens, with cached reads dropping to a fraction of a cent. For context, a million tokens is a lot of text, so those numbers are strikingly low compared with premium closed models. One independent review noted that DeepSeek positioned V4 against Gemini 3.1 Pro and GPT-5.4 on reasoning and coding, while stopping short of claiming it beats the very top closed models. In other words, the pitch is not "the best in the world" so much as "remarkably close, for a tiny fraction of the cost, and you can run it yourself."

Kimi K3: the biggest open model yet

If DeepSeek's story is about efficiency, Kimi K3 from Moonshot AI is about scale. Moonshot released the full weights in late July 2026 and describes it as the world's first open model in the three-trillion-parameter class. On paper it is a Mixture-of-Experts model with 2.8 trillion total parameters and about 104 billion active per request. It also handles text and images natively rather than bolting vision on afterward, and it shares the same one-million-token context window as the DeepSeek models.

Under the hood Moonshot introduced a few new tricks, including an attention method it calls Kimi Delta Attention and a sparser expert design that picks 16 of 896 specialist experts for each token. The company claims this makes K3 roughly two and a half times more efficient to scale than its previous K2 model. The practical reason to care is that K3 was clearly built for long, hands-off work: writing and debugging code across big repositories, doing deep research, and driving software tools over extended sessions.

The benchmark numbers Moonshot published are genuinely eye-catching, and this is where the "wildly acclaimed" reputation comes from. On its own evaluation table, K3 posts leading or near-leading scores on several coding and agentic tests, and it edges out or matches strong closed models on a number of them. It is worth a note of caution, though: these figures come from the model maker itself, and independent testers have reported a more mixed picture on some tasks. The honest summary is that K3 is one of the strongest open models ever released, and probably not a clean sweep over the best closed systems on every measure. Its license is also its own "Kimi K3 License" rather than a standard permissive one, so anyone planning commercial use should read the terms carefully.

One more practical wrinkle: sheer size has a cost. Community reports put a full download of K3 well over a terabyte, and running it locally is out of reach for ordinary hardware. Most people will use it through an API instead, where Moonshot's pricing lands around $3 per million input tokens and $15 per million output tokens, considerably higher than DeepSeek Flash though still competitive with premium closed options.

The rest of the open field: GLM, Qwen, and Llama

DeepSeek and Kimi are not the only players worth knowing. GLM-5.2, from the Chinese lab Z.ai, arrived in June 2026 under an MIT license and has become a favorite in independent coding roundups. It is a Mixture-of-Experts model with around 744 billion total parameters and roughly 40 billion active, plus the now-familiar one-million-token context window. Several reviewers rate it as the strongest open model for long-running engineering agents, though self-hosting it still demands serious server hardware.

Alibaba's Qwen family plays a different game. Instead of chasing the largest possible model, Qwen ships a wide range of sizes, including compact versions small enough to run on a laptop or a Mac with Apple Silicon. That flexibility, combined with permissive licensing, makes Qwen the go-to for people who want something they can actually run at home. Meta's Llama series, meanwhile, remains the ecosystem default for many businesses: its dense models are demanding at the top end, but the tooling, tutorials, and community support around Llama are unusually mature.

How they actually differ

Strip away the branding and a few clear differences emerge. The first is size and cost. Kimi K3 is the heavyweight, built for maximum capability on long, complex jobs, and it charges accordingly. DeepSeek V4 Flash sits at the opposite end, prioritizing speed and rock-bottom pricing while staying surprisingly capable. GLM-5.2 and DeepSeek V4 Pro occupy the middle, offering frontier-level quality without quite the scale of K3.

The second difference is what you can run yourself. Almost all of these top-tier models need heavy hardware to host, so the realistic "run it on your own machine" option usually comes from smaller Qwen variants rather than the giants. The third is modality: Kimi K3 natively understands images as well as text, whereas the DeepSeek V4 models are primarily text-focused in their standard form. The fourth is licensing, which is easy to overlook. DeepSeek V4 and GLM-5.2 use the well-understood MIT license, while Kimi K3 ships under its own custom license that is worth reading before you build a business on it.

Quick pros and cons

DeepSeek V4 Flash. Pros: extremely cheap, fast, easy to justify at scale, and strong on coding for its size. Cons: not the outright quality leader, and mainly text-focused.

DeepSeek V4 Pro. Pros: frontier-level quality, permissive MIT license, huge context. Cons: large and demanding to self-host.

Kimi K3. Pros: arguably the most capable open model to date, native vision, built for long agentic work. Cons: enormous and impractical to run locally, pricier API, and a custom license.

GLM-5.2. Pros: top marks in independent coding tests, MIT license. Cons: still needs serious hardware to self-host.

Qwen and Llama. Pros: Qwen offers the best small, laptop-friendly options; Llama has the most mature ecosystem. Cons: neither top tier trades blows with the very largest new models on the hardest tasks.

So which one should you care about?

For most everyday needs, the smart move is to match the model to the job rather than chase the biggest name. If you want the cheapest way to handle a lot of text or code, DeepSeek V4 Flash is hard to beat. If you want the most powerful open model for long, complex, tool-heavy work and do not mind paying more, Kimi K3 is the flagship. If coding agents are your focus, GLM-5.2 and DeepSeek V4 Pro both deserve a look. And if you genuinely want to run something yourself on modest hardware, a smaller Qwen model is the practical choice.

The bigger picture is the real headline. A year ago, the best models were almost all locked behind closed doors. Now open releases are trading blows with them, often at a fraction of the cost, and new ones arrive nearly every month. Whichever model wins your particular use case, the trend itself is the thing to watch.

Sources and further reading

DeepSeek V4 Flash 0731 model card and benchmark table: huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

DeepSeek V4 preview release notes and pricing: api-docs.deepseek.com news, April 24, 2026

Kimi K3 model summary, architecture, and evaluations: github.com/MoonshotAI/Kimi-K3 and the Kimi K3 tech blog at kimi.com

GLM-5.2 specifications and coding-model roundups: independent reviews at howaiworks.ai and faros.ai, June to July 2026

Note: several benchmark figures cited above come directly from the model makers, so treat vendor-reported scores with healthy skepticism and cross-check against independent testing where possible.