The short answer
DeepSeek hasn't confirmed these are V4. What we have is a large, unusually consistent wave of suspected gray-test output: 121 Bilibili videos from 53 creators between Jul 7 and Jul 19, 2026, plus 40 shared OpenCode conversations and a handful of downloadable projects. That's enough to spot patterns — not enough to certify a model ID or run a controlled benchmark.
Read the evidence that way and the picture is clear: V4 looks slightly behind Kimi K3, GPT-5.6 and Fable 5.0 in their strongest pockets, but close enough that, for most daily coding, a cost-conscious developer could default to V4 and keep one light premium subscription for review, nasty bugs, and the last 10% of long multi-turn work.
For the screenshot gallery and chat links from the earlier wave, see our DeepSeek V4 gray-release evidence roundup.
Full source catalog Key evidence
The raw evidence behind every claim in this article — mirrored from the open YunhaoFu/dsv4ga-news-gather repository. Expand to browse by creator, date or keyword. Titles link straight to Bilibili.
Click to expand all 136 records
Tip: the title is the link to Bilibili. green = shared chat session, orange = project download. Hover a creator to see the full name.
What 121 tests actually show
The data comes from the open dsv4ga-news-gather repository, which indexes public Bilibili posts and preserves titles, creator names, timestamps, descriptions, comments, shared conversations and download links. The Jul 19 snapshot breaks down like this:
Posts per day
Slow trickle until Jul 15, then a wall of releases peaking Jul 18.
Top 10 creators
Only kdzzzds and Delight-linger shared substantial session logs. Green = 5+ shared sessions, orange = 1–4, blue = none.
What people actually tested
Categorised from 136 titles. Games dominate — production backends, security and long-horizon work barely appear.
Two biases matter more than any single number. Selection bias: most of these are "wish coding" tasks — browser games, 3D scenes, SVG animation, physics toys, music tools. The set is strong evidence for rapid prototyping and front-end generation; it's weak evidence for production backends, security, large-repo maintenance, or anything that takes months. Transparency bias: only 3 of the top 10 creators shared their prompts or chat logs. Treat video titles that say "official version" as marketing — the creator usually only suspects a gray route. We use suspected V4 gray build throughout.
What V4 actually seems good at
One-prompt apps now run, instead of just looking pretty
The strongest pattern is broad implementation from a thin spec. The set includes voxel worlds, shooters, rhythm games, survival games, three.js scenes, music generators and engineering visualisations. A representative GTA-style SVG demo was described as one main generation plus one bug-fix round, with a shared conversation attached. That's not "one sentence, one shipped game". It's "the model picks an architecture, wires enough of it together to demo the idea, and no longer hands back an empty mockup".
Strong 3D, SVG and lightweight simulation
Three.js scenes, custom SVG assets, game physics and interactive simulations repeat across the collection. An air-liquid CFD-style page reportedly used PBF for liquid and LBM for gas in two conversations. The creator also flagged unrealistic aerodynamics and the need for more repair — exactly the line between an impressive prototype and a trustworthy engineering tool.
Price may be the real headline
Community comparisons repeatedly describe V4 and Kimi K3 as close on practical projects. If the final API keeps DeepSeek's usual price advantage, "near-frontier capability at infrastructure pricing" matters more than winning any single head-to-head demo. See our AI API price comparison for current context.
Where V4 still struggles
- Polish and taste. Kimi K3 is often preferred for front-end composition and long-form wording. V4's visuals can be impressive, but quality swings sharply with prompt design.
- Instruction boundaries. One demo asked the model to delete comments; it also fixed a game bug on its own. That initiative feels magical in a toy, and is a review risk in a real repo. Review every diff — don't accept "it runs" as sufficient.
- Long multi-turn reliability. Sessions still crash, lose direction, or accumulate architectural debt over many rounds. A strong first pass doesn't guarantee a strong twentieth pass.
- Physics correctness. A convincing fluid or gravity simulation is not valid CFD or mechanics. Visually right ≠ numerically right.
- Reproducibility. Gray routing looks inconsistent. Different users may have hit different models, tiers or sampling behaviour.
- Evidence quality. Many videos omit the full prompt, model identifier, failure count, token use and manual edits.
The honest description is "high-ceiling prototype generator with real coding ability" — not "a senior engineer that no longer needs review".
V4 vs Kimi K3 vs GPT-5.6 vs Fable 5.0
There's no controlled four-model benchmark here. The table below is a workflow-oriented synthesis of the gray-test evidence and the user reports around it — not a leaderboard.
| Model | Likely advantage | Where V4 stands | Best role |
|---|---|---|---|
| DeepSeek V4 | Cost, broad implementation, rapid prototypes | Baseline | Default daily coding and first implementation |
| Kimi K3 | Front-end polish, presentation, writing | V4 is slightly less polished but often functionally comparable | UI pass, product copy, visual refinement |
| GPT-5.6 | Review consistency, tool ecosystem, cross-file reasoning | V4 may be the cheaper substitute for routine work; not enough evidence for a general win | Code review, validation, stubborn bugs |
| Fable 5.0 | Long-horizon implementation, multi-turn recovery | V4 looks close in selected demos but less reliable at the edge | Escalation model for the hardest remaining work |
V4 vs Kimi K3: the most credible direct material points to a close contest. K3 looks better on front-end and writing; V4 may be stronger value and competitive on implementation. One comparison video reached tens of thousands of views, but prompts and scoring weren't standardised.
V4 vs GPT-5.6: isolated comments claim wins and losses on particular web projects. These are test ideas, not benchmark results. The useful question is whether V4 can pass your repo's test suite at lower cost.
V4 vs Fable 5.0: impressive game demos make the comparison tempting, but the collection doesn't establish parity on large codebases, security review, or long autonomous runs. Treat Fable 5.0 as an escalation option until reproducible tests say otherwise.
A practical setup
The cheapest setup that doesn't bite you later:
- Let V4 do the first 80–90%. Planning, scaffolding, implementation, tests, docs, ordinary bug fixing.
- Verify locally. Lint, type-check, unit and integration tests, and a human diff review. Non-negotiable.
- Escalate with evidence. When stuck, send the diff, failing tests and a narrow question to Kimi K3, GPT-5.6 or Fable 5.0 — not the same vague prompt again.
- Keep one light premium plan, not several full subscriptions. Use it as a reviewer and recovery model, not the default token burner.
This only works if the second model gets evidence — diffs, logs, failing tests. Asking two models the same vague question usually doubles cost without improving reliability.
What to verify when V4 officially launches
Judge the official release on reproducibility, not launch-day demos. Check:
- the exact API model ID, and whether web, app and API use the same tier;
- context window, output limit, tool calling and structured-output support;
- price per input, cached input and output token;
- rate limits, peak-hour routing, and whether "Pro/Flash/Max" tiers behave differently;
- repository-scale editing, test completion and instruction adherence;
- how often identical prompts reproduce the gray-test quality;
- licence, data retention and deployment options.
We'll update this page when DeepSeek publishes official model cards, API docs and pricing. Until then, claims about the final model stay provisional.
FAQ
Is the official version already available?
The collection documents suspected gray routing, not a confirmed release. A video titled "official version" is not official confirmation from DeepSeek.
How good is V4 for coding?
The evidence is strongest for rapid web prototypes, games, SVG, three.js and iterative bug fixing. It's thinnest for large production repos, security-sensitive code and long autonomous work.
Is V4 better than Kimi K3?
Not across every task. V4 appears close and may offer better value; Kimi K3 often has an edge in front-end polish and writing. Run the same prompt, runtime and acceptance tests before choosing.
Should I cancel my premium coding subscription?
For many users, downgrading to one lightweight premium plan makes more sense than cancelling everything. Let V4 handle volume and keep an independent model for review and escalation.
Where can I inspect the original gray tests?
Start with the GitHub index of all 121 videos. Representative cases: broad V4 + Kimi K3 test, 3D physics comparison, 8192-block Minecraft test, proactive bug-fix example.