Direct Answer
DeepSeek V4 Flash 0731 is one of the most interesting low-cost coding models available, but “small” needs an asterisk. The official model has 284 billion total parameters, activates about 13 billion per token, supports a one-million-token context window, and publishes its weights under MIT. In Bijan Bowen's eight ambitious builds, it produced unusually complete games, interfaces, music, and CAD files for a creator-reported $0.79 of DeepSeek API spend.
The same test also found a stable weakness profile: inverted controls, wrong 3D orientation, mangled vehicles, fragile game state, missing enemies, and generated objects that looked plausible before close inspection. This is a strong prototyping worker. It is not a one-prompt production engineer, a reliable CAD authority, or a model to run unattended against important systems.
Watch the Full Test
Video credit: Bijan Bowen. Watch the original video on YouTube. Model architecture, license, benchmark settings, API availability, and pricing were checked against DeepSeek's official model card and API documentation on 3 August 2026. Visual quality judgments and the $0.79 spend are creator observations from this run.
What DeepSeek Actually Released
DeepSeek-V4-Flash-0731 is the official release that supersedes the preview. DeepSeek says the architecture and size are unchanged and that the improvement comes from additional post-training. The release also attaches DeepSeek's DSpark speculative-decoding module, supports low, high, and max reasoning effort, and is adapted to the Responses API used by Codex.
| Specification | Official value | What it means |
|---|---|---|
| Architecture | Mixture of Experts | Only part of the network is active for each token. |
| Total parameters | 284B | Weights and memory remain very large even though active compute is lower. |
| Activated parameters | 13B | The source of the “small” and low-cost framing. |
| Context | 1M tokens | Useful for large repositories and long agent traces; full context still needs substantial KV-cache memory. |
| Maximum output | 384K tokens | A ceiling, not a sensible default budget. |
| Reasoning effort | Low, high, max | Bijan tested at max, favoring capability over speed and cost. |
| Weights | MIT License | Commercial reuse and modification are permitted under the license terms. |
| Hosted access | Public beta API | Use deepseek-v4-flash for the current 0731 endpoint. |
Is It Really a Small Model?
Calling a 284B model “small” is reasonable only inside the narrow world of frontier Mixture-of-Experts systems. V4 Flash activates 13B parameters for a token, far below DeepSeek V4 Pro's 49B active path and Kimi K3's much larger total footprint. That reduces inference work. It does not shrink every weight, the context cache, runtime buffers, or the speculative module into a normal laptop package.
A better description is small active path, large physical model. A 13B dense model may fit comfortably on a consumer machine. A 284B MoE still needs all experts available in memory or across devices, even though only selected experts execute for each token. The distinction matters before anyone buys hardware based on the headline.
The Real API Price
The video reads the cache-hit figure as roughly three cents. DeepSeek's pricing table actually lists $0.0028 per million cache-hit input tokens: twenty-eight hundredths of one cent. The cache-miss and output figures in the video are correct. DeepSeek also announces a future peak-hours multiplier, but the same page says the effective date will follow in a separate announcement.
| Billing item | Regular price per 1M tokens | Future peak-hours price |
|---|---|---|
| Input, cache hit | $0.0028 | $0.0056 |
| Input, cache miss | $0.14 | $0.28 |
| Output | $0.28 | $0.56 |
During the video, the displayed DeepSeek balance moves from $9.85 to $9.06, implying $0.79 for the model work shown. The Browser OS alone moves the balance from $9.85 to $9.77, about eight cents. These numbers are useful directional evidence, not a controlled cost benchmark: tasks ran with different lengths, one game ran concurrently, cache status was not reported, and the Steve's PC Repair build used a separate OpenAI API key for dialogue.
The practical buying metric is not dollars per million tokens. It is cost per accepted result. A $0.10 build that needs an hour of debugging may be more expensive than a $2 build that passes its tests. Track model spend, human repair time, retries, and verifier cost together.
The Eight-Test Scorecard
| Test | Strongest signal | Observed failure | Evidence confidence |
|---|---|---|---|
| Browser OS | Resizable windows, files, right-click menu, personalization, games, snapshots, local date and time | Calculator error, incomplete settings, inverted or mismapped game controls | Medium: rich result, but the prompt is a recurring public benchmark |
| C++ skate game | Complete retro scene, pedestrians, tricks, rail balance, beach and water interaction | Broken text and tricks on first pass; awkward HUD after repair | Medium-high: improved through visible iteration |
| Subway FPS | Atmosphere, sound, bloom toggle, flashlight, bullet impacts, exploration and combat modes | No enemies initially, auto-firing weapon, unreliable wave logic | High for visual prototyping; low for game correctness |
| C++ MotoGP | Strong retro racing presentation in both runs | First run violated the no-raylib constraint; second may have retained prior context; controls appeared inverted | Low: Bijan correctly calls the test ruined |
| Luxury watch site | Good materials, terminology, rendered assets, sub-dials, and product-page structure | Watch parts were arranged or oriented incorrectly in 3D | Medium: good style signal, poor product fidelity |
| Printable inline-six shell | Multiple STL parts, shaft opening, vents, support-aware geometry, motor fit-check model | Motor orientation wrong, some walls too thin, uncertain ring and cover fit | Medium: promising CAD draft, no physical print test |
| C++ Chrono City | Five eras, changing architecture, signs, palette, music, people, and cinematic transitions without raylib | Long black-screen debugging, mangled cars, odd movement, hard-to-control tour | Medium-high for ambition; medium for usability |
| Steve's PC Repair | Long narrative, repair interaction, dialogue, inventory, travel, stealth scene, and multi-stage progression | Unclear navigation, pacing too fast, spatial confusion, progression became frustrating | Medium: impressive scope, incomplete playability |
What V4 Flash Did Well
The strongest pattern was not perfect code. It was scope completion. V4 Flash repeatedly turned broad prompts into working folders with application structure, rendering, sound, assets, interactions, and an attempt at verification. That is valuable for builders who need ten plausible directions before choosing one.
- End-to-end prototyping: it built playable scenes rather than stopping at scaffolding or an implementation plan.
- Multimodal-feeling craft from text: materials, music, lighting, UI, and motion often formed a coherent mood even when geometry was wrong.
- Iteration after feedback: the skate game recovered an ollie, tricks, and some text after a short correction loop.
- Tool endurance: the model kept working through compile, browser, rendering, and file-generation tasks under max reasoning.
- Novel combinations: C++ Chrono City and the inline-six motor shell pushed beyond familiar landing-page tests.
- Economics: the reported endpoint spend is low enough to run several candidates and spend the saved budget on verification.
Where It Failed, Repeatedly
A good evaluation looks for recurring failure modes, not a gallery of best screenshots. Across unrelated prompts, V4 Flash showed the same weaknesses often enough to treat them as routing information.
- Spatial reasoning: 3D watch parts, motor placement, vehicle orientation, and camera controls were often wrong while still looking superficially polished.
- Control mapping: inverted movement and unclear key bindings appeared in the Browser OS games, MotoGP, Chrono City, and PC Repair.
- State and game logic: missing enemies, wave zero, auto-fire, unclear task progression, and over-fast dialogue reduced playability.
- Constraint obedience: the first MotoGP run used raylib despite an explicit prohibition.
- Verification depth: rendered screenshots and plausible geometry are not proof that a part fits, a game loop completes, or a calculation is correct.
Benchmark Context: Strong, Not an Overall Winner
DeepSeek reports a large jump from the preview after post-training. On the public tests listed in the official release, V4 Flash 0731 scores 82.7 on Terminal-Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. It beats DeepSeek V4 Pro Preview across DeepSeek's table, despite the smaller active path.
It does not beat Opus 4.8 across that same table. Opus is higher on every listed public benchmark: 85.0 on Terminal-Bench, 69.7 on NL2Repo, 83.1 on Cybergym, 58.0 on DeepSWE, and 76.2 on Toolathlon-Verified. DeepSeek's two DSBench results are internal. The code-agent runs also use a forthcoming DeepSeek Harness in minimal mode, max effort, temperature=1.0, and top_p=0.95.
That makes “best small model” a reasonable creator verdict, not a settled benchmark fact. A useful comparison would run V4 Flash, Kimi K3, GPT-5.6, and Claude on the same repository snapshot, same acceptance tests, same tool permissions, same retry budget, and fresh sessions. Bijan's video is a strong capability probe; it is not that controlled experiment.
Local Hardware Reality
DeepSeek's official mixed-precision checkpoint uses FP4 expert weights and FP8 for most other parameters. Its published vLLM example serves the 0731 release on a single four-GPU GB300 node. That is the clean reference deployment, and it is data-center hardware.
The video points to a community four-bit quant around 138 GB and more aggressive dynamic quants that might squeeze into a 128 GB unified-memory machine. “Might fit” needs five caveats: the quant is not the official checkpoint format, quality may fall, runtime overhead also needs memory, a long context can make the KV cache enormous, and token speed may be too slow for interactive agent work. Two DGX Spark systems are an enthusiast route, not a consumer baseline or an official requirement.
| Route | Best for | Main tradeoff |
|---|---|---|
| DeepSeek API | Most developers, evaluation, bursty agent work | Hosted-data policy, network dependence, changing prices |
| Third-party hosted endpoint | Region or infrastructure choice | Provider quality, terms, and model-version drift |
| Official mixed-precision self-host | Labs and companies with multi-GPU infrastructure | Hardware, power, orchestration, and operating cost |
| Community low-bit quant | Local AI enthusiasts exploring fit and privacy | Quality, speed, context, tooling, and reproducibility |
| Smaller dense local model | Daily private work on ordinary hardware | Lower peak capability, often much better responsiveness |
Using DeepSeek V4 Flash in Codex
DeepSeek publishes an official Codex integration guide. V4 Flash supports the Responses API, and DeepSeek says one Codex configuration can expose the model to the CLI, desktop app, and IDE extension. The guide offers an automated setup script and a manual configuration path.
Prefer the manual path, or download and inspect the script before execution. A command that pipes a remote script directly into a shell gives the publisher immediate code execution on your machine. The official script says it backs up ~/.codex/config.toml, writes a model catalog, preserves MCP and trust settings, and validates files before writing; inspecting the current script is still the safer habit.
- Create a separate DeepSeek API key with a small prepaid balance.
- Back up the Codex configuration and inspect the provider changes.
- Open a fresh, non-sensitive repository with no production credentials.
- Start at
higheffort; reservemaxfor genuinely difficult tasks. - Give the task a written acceptance test and a hard cost or time limit.
- Review the diff and run the repository's full verification before merging.
A Better Seven-Step Evaluation
- Choose one real task. Use a bug, feature, migration, or design repair from your own work rather than a famous public benchmark.
- Freeze the environment. Same commit, dependencies, tools, permissions, and machine for every model.
- Write acceptance criteria first. Include tests, screenshot states, performance limits, accessibility checks, or dimensional tolerances.
- Start fresh sessions. This prevents prior attempts from leaking hints into later runs.
- Cap retries. Give each model the same time, tokens, and number of repair loops.
- Score the whole outcome. Track pass rate, latency, API cost, human repair minutes, and regressions.
- Route by failure profile. Keep V4 Flash for the tasks where it wins on accepted-result cost, not where it only makes the best first screenshot.
Data and Privacy Before Hosted Use
DeepSeek's current public privacy policy says its services may collect prompts, uploaded files, feedback, chat history, and generated outputs; may use data to improve and train its technology; offers a training opt-out; and processes and stores personal data in the People's Republic of China. It also says the services are not designed for sensitive personal data.
That does not automatically disqualify the API. It does mean a cheap model should not bypass normal vendor review. Keep secrets out of prompts, redact client data, use synthetic fixtures, restrict agent permissions, rotate keys, and obtain legal or security approval before sending confidential repositories. Self-hosting the MIT weights can change the data path, but then patching, access control, logging, and infrastructure security become your responsibility.
Claim Audit
| Claim | Assessment | Context |
|---|---|---|
| V4 Flash is the best small model yet. | Opinion, plausible | Excellent creator results and low cost; no controlled cross-model test establishes a universal winner. |
| The model is 284B total and 13B active. | Confirmed | Published in DeepSeek's official V4 model card; the 0731 release keeps the same architecture. |
| It has a one-million-token context. | Confirmed | Listed in the official model card and API pricing documentation. |
| Cache hits cost about three cents per million input tokens. | Incorrect | The official figure is $0.0028, about 0.28 cents per million. |
| All shown tests cost $0.79. | Creator-reported | The displayed DeepSeek balance falls by $0.79; a separate OpenAI dialogue cost and human time are excluded. |
| It can run on one 128 GB machine. | Possible only with caveats | An aggressive community quant may fit, but speed, overhead, context, quality, and runtime support determine practicality. |
| V4 Flash 0731 beats V4 Pro Preview. | Project-reported | It leads across DeepSeek's published table after additional post-training. |
| The future V4 Pro will beat Kimi K3. | Speculation | Flash improvements do not guarantee how the unreleased Pro update will perform. |
| The model is safe to ship from one prompt. | False | The tests show repeated logic, geometry, control, and constraint failures despite polished presentation. |
Video Chapters
| Time | Chapter | Evaluation signal |
|---|---|---|
| 00:00 | Intro | Official 0731 release and the “small model” thesis. |
| 01:11 | Local AI info | 284B total, 13B active, community quantization estimates. |
| 02:45 | Technical look | Benchmark jump, API access, context, and pricing. |
| 04:26 | Browser OS test | Rich interactivity plus arithmetic and control failures. |
| 11:11 | C++ skate game | Strong scene construction and useful iteration after feedback. |
| 14:13 | Subway FPS | Excellent atmosphere; missing enemies and broken combat state. |
| 16:50 | C++ MotoGP | Visually strong but methodologically contaminated. |
| 18:52 | Watch website | Good materials and copy; weak 3D orientation. |
| 20:47 | 3D CAD model | Printable-looking parts with dimensional and fit concerns. |
| 23:35 | C++ Chrono City | Novel multi-era build with music, motion, and spatial defects. |
| 28:04 | Total API cost | Creator-reported $0.79 DeepSeek balance change. |
| 28:23 | Steve's PC Repair | Long-form narrative ambition, but unreliable navigation and pacing. |
| 34:04 | Results overview | Strength and failure patterns across all tests. |
| 36:23 | Closing thoughts | Local-hosting enthusiasm and unverified V4 Pro forecast. |
Bottom Line
DeepSeek V4 Flash 0731 is not impressive because it made one pretty webpage. It is impressive because a 13B-active open-weight model repeatedly assembled ambitious, multi-file, interactive projects at a remarkably low hosted price. Its official agent benchmarks also show that post-training moved it far beyond the preview.
The video is equally useful as a warning. Polished outputs concealed arithmetic mistakes, wrong orientations, violated constraints, broken controls, and incomplete state logic. Those are exactly the errors that become expensive when a prototype is mistaken for a finished system.
The sensible adoption path is simple: test the API on a real low-risk task, give every task an objective verifier, measure accepted-result cost, and route work by observed strengths. Keep local V4 Flash as an infrastructure decision, not an impulse hardware purchase. Under those conditions, this model can be a very capable and unusually affordable worker.
Sources and Useful Links
- Bijan Bowen: DeepSeek V4 Flash Is INSANE - The Best Small Model Yet
- Bijan Bowen on YouTube
- DeepSeek-V4-Flash-0731 official model card and MIT weights
- DeepSeek V4 architecture, model sizes, and technical report links
- DeepSeek API change log: 31 July V4 Flash update
- DeepSeek models and current API pricing
- Official DeepSeek integration guide for Codex
- DeepSeek concurrency limits and user isolation
- DeepSeek privacy policy
- DeepSeek Open Platform terms of service