📖 Table of Contents
1. TL;DR — The Verdict
DeepSeek V4 Flash is the official "fast and cheap" sibling of DeepSeek V4 Pro, built on a Mixture-of-Experts architecture with 284B total parameters but only 13B active per token. On the Artificial Analysis intelligence index it ranks #3 of 101 models, while its cache-hit price is the lowest in the market — a combination no other frontier lab currently matches.
Bottom line: If you need frontier-level reasoning with a hard cost ceiling — long-context RAG, coding agents, bulk classification, or multilingual workloads — V4 Flash is the most interesting price-performance option released this year. If you need the absolute best single-shot reasoning on open-ended research, the top commercial models still hold the edge.
2. Specs: What "284B / 13B Active" Actually Means
| Property | DeepSeek V4 Flash |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Total parameters | 284B |
| Active parameters per token | 13B |
| Context window | 1M tokens |
| License | MIT (open weights) |
| Intelligence index (Artificial Analysis) | #3 / 101 |
| Cache price (per 1M tokens) | $0.003 — lowest in market |
The key architectural idea is MoE with a tiny active set: the model has 284B parameters in total, but for each token it only activates 13B. This is what lets it run fast and cheap while retaining most of the knowledge capacity of a much larger dense model. The 1M-token context — up from 128K in V3-era models — is the other headline spec, making it practical for whole-repository code analysis and very long document workloads.
The MIT license is a meaningful choice: unlike some competitors that open the weights but restrict commercial use or require attribution clauses, MIT means you can fine-tune, sell, and deploy without legal friction — a big deal for startups building products on top.
3. Benchmarks: Artificial Analysis Intelligence Index
Artificial Analysis is the de-facto independent benchmark aggregator for LLMs, and its intelligence index blends reasoning, math, coding, and knowledge tasks into a single score. V4 Flash's #3 ranking puts it ahead of several much more expensive frontier models, behind only the top reasoning models from the biggest labs.
What's remarkable is not the absolute score — it's the score per dollar. In the same release window:
- Models ranked 1-2 cost 10-50× more per token at comparable context lengths.
- V4 Flash delivers ~90% of the top model's benchmark score at a fraction of the operating cost.
- On coding benchmarks specifically (SWE-bench style agentic tasks), V4 Flash is competitive with models that cost 20× more for long-context runs.
Caveat: benchmark indices measure average capability, not worst-case reliability. For high-stakes single-shot reasoning — legal analysis, medical triage, architectural decisions — the top dense reasoning models remain more predictable. V4 Flash's edge is throughput and cost at scale, not maximum depth on every problem.
ARC-AGI: 89% on Semi-Private Tasks at $0.02 per Task
Updated Aug 8, 2026: On August 7, ARC Prize published official results for the V4 Flash 0731 reasoning variant, and they're remarkable for a model at this price point:
- ARC-AGI-1 Semi-Private: 89.0% at max effort — at a cost of $0.02 per task.
- ARC-AGI-2 Semi-Private: 61.4% at max effort — $0.04 per task.
- High-effort variant: 87.0% (ARC-AGI-1) / 56.0% (ARC-AGI-2). Low-effort: 84.0% / 46.0%.
ARC-AGI-1 is the abstract-reasoning benchmark designed to resist memorization — an 89% Semi-Private score puts V4 Flash in the same band as the frontier reasoning models that cost far more to serve, while the per-task economics are a rounding error. ARC-AGI-2 (the harder, 2025-era follow-up) at 61.4% is also competitive for the price. The pattern is the same one we keep seeing: V4 Flash doesn't win every benchmark, it wins score-per-dollar — and on a pure reasoning-per-cent basis, nothing else in the market touches it right now.
4. Pricing: The Cache Price Disruption
This is the number that broke the Chinese developer community's calculators: $0.003 per 1M cached tokens — a ~98% discount versus typical cache pricing from US providers. For RAG-heavy or agentic workloads where the same context (a repo, a knowledge base) is re-read constantly, cache hits dominate the bill, and that math changes everything.
Practical example: a coding agent that re-reads a 200K-token repository on every turn. At V4 Flash's cache price, a 100-turn session costs a fraction of a cent in context reads. The same session on a US frontier model with similar context would run to dollars. For a team running thousands of agent sessions a day, this is the difference between "experiment" and "product."
Who should care: (1) RAG/agent startups with heavy context reuse, (2) companies doing bulk multilingual processing, (3) developers self-hosting via open weights on their own GPU clusters, (4) anyone whose current LLM bill is dominated by input tokens rather than output tokens.
5. Real-World Feedback from V2EX
V2EX — the Chinese developer community — had three active threads about the V4 Flash release within hours, and the sentiment is genuinely mixed, which is itself informative:
What users love
- Price: unanimous agreement that cache pricing and output pricing reset expectations for what a "cheap" frontier model costs.
- Long context: 1M tokens works well for whole-repo analysis; several users reported feeding entire codebases without chunking.
- Chinese-language quality: expected given DeepSeek's heritage — users report native-level Chinese output that beats US models on Chinese idioms, nuance, and coding comments.
What users criticize
- V4 Pro value question: several users argue that if Flash is this good, V4 Pro's premium pricing is hard to justify — a "cannibalization" complaint that is actually a compliment to Flash.
- Reasoning depth on hard problems: math competitions and adversarial logic puzzles occasionally expose shallower reasoning than the top dense models.
- Tool-calling edge cases: some agent harnesses report occasional malformed tool calls in complex multi-step workflows — a maturity gap versus providers with years of agentic tuning.
6. V4 Flash vs V4 Pro vs GPT-5 vs Claude Opus
| Criterion | V4 Flash | V4 Pro | GPT-5-class | Claude Opus-class |
|---|---|---|---|---|
| Intelligence (index) | #3 / 101 | #1-2 | #1-2 | Top 5 |
| Context | 1M | 1M | 128K-1M | 200K-500K |
| License | MIT open | MIT open | Closed | Closed |
| Cache price /1M | $0.003 | ~10-30× Flash | $0.10-0.50 | $0.10-0.50 |
| Best for | Scale + cost | Max reasoning | Ecosystem + tooling | Coding + safety |
Positioning summary: V4 Pro is the "think hardest" model. V4 Flash is the "think well, at scale, for cheap" model. GPT and Claude retain advantages in mature agentic tooling, enterprise ecosystems, and long track records of reliability — but they are priced 1-2 orders of magnitude higher for comparable workloads.
7. Pros & Cons
Strengths
- Top-3 intelligence at a fraction of the price
- Lowest cache price in the market ($0.003/1M)
- 1M token context — whole-repo and whole-book workloads
- MIT license: commercial use, fine-tuning, self-hosting all free
- Excellent Chinese-language output
- Fast inference thanks to 13B active params
Limitations
- Occasional shallow reasoning on hard single-shot problems
- Agentic tool-calling less mature than US providers
- Ecosystem (plugins, enterprise support) still developing
- Self-hosting 284B params requires serious GPU infrastructure
- V4 Pro's value proposition now looks squeezed
8. FAQ
Q: Is DeepSeek V4 Flash really open source?
The weights are released under the MIT license — one of the most permissive licenses available. You can use it commercially, modify it, and redistribute it. "Open source" in the strict OSI sense is debated (training data is not released), but for practical purposes it is open-weight with no commercial restrictions.
Q: Can I run V4 Flash on my own hardware?
Yes, but note the scale: 284B total parameters. With MoE, only 13B activate per token, so inference is feasible on a single high-end GPU (e.g., 80GB-class) with aggressive quantization, but full-precision serving realistically wants a small cluster. Most users will start with the hosted API and move to self-hosting once usage justifies the hardware.
Q: What workloads benefit most from the cache price?
Anything that re-reads the same context repeatedly: RAG pipelines, coding agents working on one repository, document analysis with long templates, and multi-turn conversations with large system prompts. If your context is highly repetitive, cache hits dominate and V4 Flash's price advantage is extreme.
Q: How does it compare to DeepSeek V3-era models?
V4 Flash is a generational step: roughly 1M vs 128K context, a large jump in the intelligence index, and better tool-calling. V3 models remain usable but you're leaving capability on the table for almost no price savings.
Q: Should I switch my agent harness to V4 Flash?
For cost-sensitive, high-volume agent workloads: very likely yes. For a few high-stakes reasoning tasks where you want maximum reliability: keep a top dense model as fallback. A good pattern is routing: V4 Flash for the bulk, frontier models for the hard cases.
Q: Where can I try it?
- DeepSeek official API (api.deepseek.com) — V4 Flash endpoints live
- Artificial Analysis: full benchmark data
- Open weights on Hugging Face / ModelScope (MIT)
- ChatGPT vs Claude vs Gemini vs DeepSeek (2026) — how the families compare