Deepseek V4 then V4.1 Flash have been my workhorse models for months, both locally & on the APIs (but recently, GLM 5.3f, Mimo 2.6f, & Qwen 3.8f Next have been great alternatives for local hosting)

The new version of V4 Flash out of preview (0731) is going viral, because it scores even higher than V4 Pro on many benchmarks, and users are seeing it produce fantastic quality output. It's still the same model architecture, but they apparently improved the post-training data quality significantly (the vast set of devised question-answer pairs which train the model how to act, and sculpt the sorts of results it delivers).

The 0731 build is live on deepseek.com and Openrouter now, and the open weights have been released. The open source versions are a drop-in replacement for the preview versions, in the ds4 framework, for those who are self hosting.

The 2-8 mixed bit quant still runs well and does a very good job, for all the sorts of tasks I rely on, when served on a single machine with 128Gb VRAM. The 4 bit quant does even better job, with more room for full length KV cache, on a cluster of 2 of the same machines (DGX Sparks, Strix Halos, Macs with 128Gb, 2 RTX 6000s, etc.).

I expect Deepseek V4 Flash on 2 Asus GX10 machines to be the default foundation for self-hosted inference, which I'll recommend to clients. Qwen 3.6, Gemma 4, Mimo 2.5, and Hy3, are all still great alternatives for fast sub-task work, but V4 Flash will be my local LLM backbone. It's proven itself to be rock solid in all my work.

V4 Flash has honestly come to dominate a large portion of all my production LLM use. I've been using it to complete virtually every IT-related task, since the day I began using it, and I've never seen it fumble. It's ridiculously fast, extraordinarily capable for its size, and outrageously cheap on the Cline-pass provider, plus there are plenty of backup providers, including Openrouter, deepseek.com, Ollama, and elsewhere. I've used V4 Flash to build an absolutely huge variety of small-medium sized utility applications, and many significant portions of code bases that fit into larger production software stacks. It's been a rock solid reliable developer, in the Pi coding harness (I should make it clear that my ChatGPT zip file development routine still forms the foundation of most large-scale production work for my big clients, and it will continue to be my go-to LLM for large Python/Flask/Web apps, for as long as OpenAI provides apparently endless rate limits for frontier use, at $20 per month - but I am no longer frightened of the moment they go out of business).

V4 Flash has absolutely become my go-to model for any significant local inference work. I perform lots of tasks which involve PHI (health data that requires strict HIPAA compliance to avoid legal repercussions), which can't be entered into typical public LLM API providers such as Openrouter - or which is very expensive to process on HIPAA compliant providers such as Bastion. V4 Flash is my go-to for any challenging tasks in that domain.

I still use Qwen 3.6 often for simple local tasks that must be compliant, because it performs so much faster than V4 Flash, but if there's any question about the potential for a less than acceptable quality of work by Qwen, V4 Flash gets the job from the get-go. I wrote a bit about one such experience in this topic: https://aibynick.com/thread/60

I've gotten V4 Flash performing at 13-27 tokens per second locally, depending on which version I'm running and other environment factors. That's plenty fast for most of the batch jobs I use it to complete, so I tend to skip even trying smaller models most of the time.

I did see one reviewer claim that V4 Flash hallucinates certain classes of work more than other similarly sized models: https://www.youtube.com/watch?v=iT34JqYtGRs , so he prefers to rely on Hy3 (but that same person also made this glowing review: https://www.youtube.com/watch?v=wT42SgaOPK4 , so take his contrary results with a grain of salt).

The majority of other initial V4 Flash reviews have been dramatically good. There are so many fantastic examples of all kinds, where V4 Flash's output beats the quality of all but the very best most recent frontier models (it even beats Kimi K3 in low reasoning mode, in some benchmark results (what?!?!? holy moly that's impressive)), and its amazing benchmark improvements compared to the preview release support the likelihood that version 0731 competes with frontier models in many domains of work.

V4 Flash provides a perfect mix of capability, performance, and self-hostable size, to make it a smart fit for an enormous scope of work. It enables you to run the same model on an API for speed (virtually for free on a provider like Cline-pass), and locally for privacy and self-reliance. The consistency that comes from using the exact same model for so many tasks, in so many situations, makes a big difference in how you work. You do become accustomed to expected workflow patterns, verbiage, strengths/weaknesses, etc., with a model that you know intimately and have used repeatedly. The experiences you work through, while building solutions with the same model, form a baseline workflow which is similar to personal familiarity you gain while working with colleagues over long term development projects. That familiarity breeds trust, reliability, and improved productivity. Switching workflows and expectations, testing results, etc., takes up an enormous amount of time, when dipping your feet into production work with a new model.

I expect V4 Flash will shape the industry in practical directions which go well beyond just constantly competing to build the next bigger multi-trillion parameter frontier model. New levels of frontier intelligence will certainly continue to shape the future, but most of us just need an affordable, fast, and capable model like V4 Flash to get the overwhelming majority of daily work completed. We need more of this sort of model: smaller, faster, cheaper, smarter for its size, and genuinely usable on modest self-hosted hardware.