DeepSeek V4 Flash Is Now on Felo — Free for Pro Subscribers for 7 Days
DeepSeek V4 Flash is now on Felo. Existing Felo Pro subscribers get 7 days of free access starting August 6 — with Search, Codex/CC, and Deep Research included.
DeepSeek V4 Flash shipped its production release on July 31, and the numbers took people by surprise. The preview from April was fine. This version is different. Agent benchmarks jumped. DSpark speculative decoding pushes throughput past 60 tokens per second. And starting today, August 6, existing Felo Pro subscribers get 7 days of free access to V4 Flash — with Search, Codex, and the full premium toolkit included. No upgrade, no extra payment, no credit card.
V4 Flash was already the best price-to-performance model on the market. Now it trades blows with flagship-tier systems on the engineering tasks people actually do every day. Here is what changed, why it matters, and what this 7-day window actually includes.
What changed between April and now
V4 Flash is not a stripped-down V4 Pro. It was built for different goals: raw speed, low resource consumption, and usability in real pipelines rather than benchmark scoreboards.
The architecture: 284 billion total parameters in a Mixture of Experts layout, with only 13 billion active per token. That is less than a third of DeepSeek V3's active count. It still beats V3 across nearly every practical benchmark. The difference comes from a hybrid attention design — Compressed Sparse Attention (CSA) paired with Heavily Compressed Attention (HCA) — that cuts inference FLOPs to 10% of what V3 burned through for the same million-token context.
What that means in practice: a full 1M-token context window runs on a single 128GB machine. That is 750,000 words. An entire codebase. A 300-page legal filing. Three novels stacked end to end. No chunking, no sliding windows, no tricks.
At 1M context, V4 Flash uses roughly 7% of the KV cache memory that V3.2 needed. Not an incremental bump. A different category.
DSpark is the other piece. It is a speculative decoding module that generates and validates candidate tokens in parallel, effectively doubling throughput. The result: 60+ tokens per second on commodity hardware. Long agent sessions feel responsive instead of like waiting for a build to compile.
The model has three reasoning modes you toggle per request:
| Mode | Behavior | When to use it |
|---|---|---|
| Non-Think | Fast, no chain-of-thought overhead | Quick lookups, translations, drafts |
| Think High | Explicit step-by-step reasoning | Debugging, math, structured analysis |
| Think Max | Full reasoning depth | Hard proofs, multi-step agent workflows |
This is not cosmetic. Run Non-Think for 90% of your prompts and switch to Think Max only when a problem actually needs it. You save time and tokens without capping your ceiling.
The benchmarks that actually matter
The preview from April posted solid numbers. The 0731 production release is built around a post-training run that targeted agentic capabilities almost exclusively. The jump is not subtle:
| Benchmark | Preview (April) | Production (0731) | V4 Pro (for reference) |
|---|---|---|---|
| SWE-bench Verified | — | 79.0% | 80.6% |
| Terminal Bench 2.1 | 61.8 | 82.7 | — |
| Cybergym | — | 76.7 | — |
| DeepSWE | 7.3 | 54.4 | — |
| GPQA Diamond | — | 71.2% | — |
| LiveCodeBench | — | 55.2% | 91.6% |
SWE-bench is the one everyone quotes, and for once the fixation is justified. 79.0% puts V4 Flash within 1.6 points of V4 Pro, a model with over 5x the active parameters. It lands in the same conversation as systems that cost 10 to 50 times more per query.
The 284B-parameter Flash outperforms the April preview of the 1.6T-parameter Pro on agentic coding. Post-training is doing most of the work here, not model scale.
The honest trade-off: V4 Flash still trails Pro on raw knowledge recall and competition-grade coding problems. But those are not the tasks most engineers spend their days on. Reading PRs, debugging across files, refactoring, writing tests — on real-world engineering work, the gap is barely visible.
What the 7-day free window actually unlocks
For the next 7 days, V4 Flash is free for Felo Pro subscribers. And Pro on Felo is more than a model picker.
Pro Search with V4 Flash under the hood
Felo Search pairs V4 Flash's reasoning with live web search. Ask a question in natural language. Felo pulls from sources across languages, synthesizes what it finds, and gives you a direct answer with inline citations. No link list. No SEO landfill. Just the answer, anchored to sources you can click through and verify.
Pro Search goes deeper. More sources pulled per query. Longer contexts processed. Answers that cover edge cases instead of stopping at the obvious. Deep Research mode turns a single question into a structured multi-page report with section headers and cited sources. The kind of thing you would spend an afternoon compiling by hand.
Codex and CC integration
This is the part developers will care about. V4 Flash natively supports OpenAI's Responses API, which means it slots into Codex CLI, ChatGPT Desktop, and the VS Code Codex extension without translation layers or compatibility shims. Set deepseek-v4-flash in your config and you are running a frontier-tier coding agent.
During the 7-day window, Pro subscribers get the full setup: agentic coding with V4 Flash driving the loop, tool orchestration, multi-file edits, and a 1M-token context window that fits your entire repo without chunking.
The rest of the Pro toolkit
- 15,000 credits loaded upfront, plus 200 daily credits
- Unlimited Pro searches. Even after credits run dry, you still get 300 Pro searches per day
- File analysis: upload PDFs, CSVs, code files and have V4 Flash work through them
- Slide generation: turn search results into presentation decks
- Mind maps: visualize how sources and concepts connect
- Cross-language aggregation: Felo reads English, Chinese, Japanese, Korean sources and synthesizes in your language
No auto-renew. The free window ends after 7 days, and V4 Flash then returns to its standard Pro pricing.
How to claim the free window
Three steps, under two minutes:
Step 1: Log in to your Felo Pro account at felo.ai. This offer is for existing Pro subscribers.
Step 2: From today, August 6, DeepSeek V4 Flash is free for 7 days. The offer is live in your dashboard — no code needed.
Step 3: Open Felo Search, pick DeepSeek V4 Flash from the model dropdown, and run a Pro search. Or head to playground.felo.ai for direct chat, or point your Codex config at deepseek-v4-flash.
The 7-day window starts today, August 6. No approval queue, no payment gate, no card on file.
Why run V4 Flash through Felo instead of the API
You can call V4 Flash directly through DeepSeek's API. It is cheap. $0.14 per million input tokens, $0.28 per million output. But raw API access leaves a lot on the table.
Real-time grounding. V4 Flash is text-only. It does not search the web, and it has a training cutoff. Felo layers live search on top, so you get frontier reasoning that is anchored to current information, not stale weights.
Side-by-side model comparison. Felo's model selector includes GPT-5.5, Claude Opus 5, Gemini 3.6 Flash, and both V4 variants. Switch models mid-conversation. Start a thread with V4 Flash for speed, flip to Opus for a second opinion on tricky logic. Same context, same sources.
Search results are not dead ends. Export findings to a LiveDoc. Generate slides from key takeaways. Build a mind map to see how sources connect. The free window gives you the full pipeline.
Timed, not permanent. This is a 7-day free window for Felo Pro subscribers. When it ends, V4 Flash returns to standard Pro pricing. Use the week to decide whether it earns a permanent place in your workflow.
How V4 Flash stacks up against the other models on Felo
Felo gives you access to a stack of frontier models, not just DeepSeek. Here is where V4 Flash fits in the lineup and when you might reach for something else.
V4 Flash vs GPT-5.5. GPT-5.5 is stronger on nuanced writing tasks, creative work, and handling ambiguous instructions. But it costs significantly more per query at the API level. On Felo, both are available. If you are debugging code or running quick-turnaround searches, V4 Flash often finishes faster. If you need prose with a particular tone or highly specific formatting, GPT-5.5 is still the safer pick.
V4 Flash vs Claude Opus 5. Opus dominates on long-form writing, careful reasoning over book-length documents, and following detailed formatting instructions. V4 Flash is faster, cheaper, and now close enough on code that many developers default to it for everyday work. Think of Opus as the precision tool for when the output really needs to sing.
V4 Flash vs V4 Pro. Pro is better at knowledge-intensive tasks, competition math, and the hardest coding challenges. Flash is better at everything else, and it is fast enough that you forget you are waiting for a model. The free window covers both. Try Pro for your hardest problem and Flash for everything else.
V4 Flash vs Gemini 3.6 Flash. Both are speed-tier models with large context windows. Gemini has multimodal input (images, video). V4 Flash is text-only but has stronger agentic coding benchmarks and the MIT license advantage. If you need vision, pick Gemini. If you need an agent that writes and edits code, V4 Flash has the edge.
FAQ
Is the 7-day free window really free?
Yes — for existing Felo Pro subscribers. No credit card, no auto-renew. The window starts today, August 6, and runs for 7 days.
What happens after the 7-day window?
V4 Flash returns to standard Pro pricing. Your other Pro features are unchanged. Nothing is deleted — your conversation history, LiveDocs, and saved searches stay intact.
Do I need to install anything for Codex integration?
No. V4 Flash supports the Responses API natively. Set deepseek-v4-flash as your model in the Codex config, and it works across Codex CLI, ChatGPT Desktop, and the VS Code extension. One config change, applies everywhere.
How does this compare to using DeepSeek's own API directly?
DeepSeek's API is cheap and fast. What Felo adds is real-time web search grounding, a multi-model comparison surface, and one-click export to documents, slides, and mind maps. If you just need raw inference, use the API. If you want search-augmented answers and a workspace around the model, Felo is the better fit.
Can I use V4 Flash on mobile?
Yes. Felo Search works in any mobile browser at felo.ai. The LLM Playground at playground.felo.ai is also mobile-ready.
The first open-weight model teams actually deploy
Here is what makes V4 Flash different from other open-weight releases: engineering teams are putting it into production pipelines as a real substitute for closed-source alternatives. Not as an experiment. Not as a fallback. As the primary model.
MIT licensed. Full weights on Hugging Face. Commercial fine-tuning allowed. And with the 0731 release, the agentic capabilities crossed the line where "good enough for the price" turned into "good enough." The 7-day free window on Felo is the lowest-friction way for Pro subscribers to see if it holds up for your workflow.
Log in to Felo Pro and start using DeepSeek V4 Flash free today — the 7-day window closes on August 12.