On July 31, 2026, DeepSeek pushed DeepSeek-V4-Flash-0731 to Hugging Face and moved the official V4-Flash API into public beta. The model card is blunt about what changed: nothing architectural. Same 284B-parameter sparse mixture-of-experts. Same 13B active parameters per token. Same 1M-token context window. The entire jump comes from re-post-training.
And the jump is not small. Terminal Bench 2.1 went from 61.8 on the preview build to 82.7. DeepSWE went from 7.3 to 54.4. Cybergym nearly doubled, 38.7 to 76.7.
Price held at $0.14 per million input tokens and $0.28 per million output.
That combination — Opus-adjacent agentic scores at roughly a rounding error of frontier pricing — is the story. Not because DeepSeek won anything, but because of what it implies about where the interesting competition is happening in 2026.
The numbers, with the caveat attached first
DeepSeek's own release notes carry a warning that most coverage dropped: all benchmark numbers are vendor-reported on an unreleased harness. Run your own evals before you route production traffic. With that stated plainly, here is what the vendor claims:
| Benchmark | 0731 | Preview | Change |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | +34% |
| NL2Repo | 54.2 | 39.4 | +38% |
| Cybergym | 76.7 | 38.7 | +98% |
| DeepSWE | 54.4 | 7.3 | +645% |
| Toolathlon-Verified | 70.3 | 49.7 | +41% |
For external context: Terminal Bench 2.1 puts 0731 at 82.7 against GLM-5.2's 81.0 and Claude Opus 4.8's 85.0. Artificial Analysis places it at 50 on their Intelligence Index — third of 101 models tracked, against a median of 25 for comparable models.
The DeepSWE line is the one worth staring at. A jump from 7.3 to 54.4 on the same weights class is not a tuning nudge; it means the preview build was effectively broken at long-horizon software engineering tasks and the post-training pass fixed something structural in how the model plans and recovers. Treat single-benchmark deltas that large as a signal to verify rather than a signal to celebrate.
One more caveat from the Artificial Analysis run: this is a reasoning model, and a verbose one. It emitted 210M output tokens across the Intelligence Index evaluation against a median of 100M. At $0.28 per million that is still cheap — but your effective cost per task is output-token-driven, not sticker-price-driven. Budget accordingly.
What actually changed under the hood
The architecture is unchanged from the preview, which makes it worth restating for anyone deploying it:
- 284B total parameters, ~13B active per token. The router picks top-6 from 256 routed experts per layer, plus one shared expert.
- Hybrid attention — DeepSeek describes it as Compressed Sparse plus Heavily Compressed Attention, which is how the 1M context stays tractable.
- Native mixed precision — FP4 for the MoE experts, FP8 for dense layers.
- MIT license, ungated repo. Commercial deployment, modification, and fine-tuning permitted without a separate agreement.
- Max output: 384,000 tokens.
The MIT license on an ungated repository is the part that matters strategically. There is no request form, no acceptable-use rider, no per-seat negotiation. For a team in Tunis or Riyadh evaluating whether they can legally ship a fine-tune into a client product, that removes the single most common blocker in open-weight adoption.
DeepSeek also ships DSpark, a speculative decoding module, and reports 60–85% faster per-user generation at matched throughput. Recommended sampling for agentic use is temperature = 1.0, top_p = 0.95 — notably high, and worth respecting rather than reflexively lowering, since the post-training was tuned against it.
The API surface changed more than the model
The quieter update: v4-flash now natively supports the Responses API format and has been adapted for Codex. V4-Pro and the web models got nothing.
That is a deliberate positioning move. Instead of asking developers to rewrite their harness around a DeepSeek-shaped API, DeepSeek adapted to the shape teams already build against. If your agent loop is already written for the Responses format, the migration cost is a base URL and a key.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.deepseek.com/v1",
apiKey: process.env.DEEPSEEK_API_KEY,
});
const response = await client.responses.create({
model: "deepseek-v4-flash-0731",
input: "Audit this repo for unhandled promise rejections and open a PR.",
temperature: 1.0,
top_p: 0.95,
});Concurrency on the beta is capped at 2,500 requests. Fine for most production agent fleets, a real constraint if you are running large batch evaluation jobs.
The pricing math that actually decides things
Here is where the cache-hit number earns its place. Official 0731 pricing per million tokens:
- Cache-hit input: $0.0028
- Cache-miss input: $0.14
- Output: $0.28
Artificial Analysis ranked it first of 101 models on cache-hit pricing. That 98% discount is not a marketing garnish — for agent workloads it is the whole economic argument.
Consider a code-review agent that reads a 200k-token repository context on every invocation and emits roughly 4k tokens of review. Run it 500 times a day:
Without prefix caching: 100M input tokens at $0.14 = $14.00, plus 2M output at $0.28 = $0.56. About $14.56/day.
With prefix caching on the repo context: the bulk of that input hits cache at $0.0028 per million. The same workload lands closer to $0.85/day.
Same model, same output, roughly 17× difference — decided entirely by whether you structured your prompts so the stable context sits at the front and the variable part sits at the back. This is the single highest-leverage change most teams have not made. We covered the mechanics in depth in our guide to AI API cost optimization with prompt caching and model routing.
Self-hosting: read the memory numbers before you plan
The MIT license makes self-hosting legal. It does not make it easy.
284B parameters in the shipped mixed precision is roughly 140 GB of weights. Heavily quantized, you are still looking at 70–80 GB. DeepSeek's own guidance puts a 3-bit quantization at around 110 GB of memory, or a 4×GB300 node for full precision.
Translated: this is multi-GPU server hardware, not a workstation. The "runnable locally" enthusiasm circulating on X is true only if your definition of local includes a rack. vLLM and SGLang both support the model with MoE expert parallelism and the hybrid attention kernels, so the software path is clean — the capital path is not.
For most teams the honest answer is: use the API, and keep self-hosting in your back pocket as leverage and as a data-residency answer. The MIT license means you can move if you need to, and that optionality has value even when you never exercise it. If sovereignty is the actual driver rather than cost, our piece on digital sovereignty and sovereign cloud in Africa and MENA works through that decision properly.
Why the budget tier is the interesting race now
The frontier tier is still moving — Opus 4.8 sits above 0731 on Terminal Bench, and GPT-5.6 recently cut prices sharply of its own accord. But the gap between the best model and a model costing a fiftieth as much has compressed faster over the last year than the frontier has advanced.
For anyone building agents, that changes the shape of the design problem. When the cheap tier was genuinely weak, architecture meant picking the strongest model you could afford and minimizing calls. When the cheap tier lands within a few points of frontier on your actual task, architecture means routing: a capable, cheap model handling the high-volume path, and a frontier model reserved for the small fraction of steps that genuinely need it.
A practical shape that works today:
- Benchmark on your own task, not theirs. Take 50 real traces from your production agent. Run them through 0731 and your current model. Compare completion rate and total cost, not benchmark scores.
- Route by step, not by workload. File reads, test runs, lint fixes, and tool dispatch go to Flash. Architectural decisions and ambiguous requirement resolution escalate.
- Structure prompts for cache hits. Stable context first, variable context last. This is worth more than the model choice.
- Keep a fallback. Beta APIs have beta reliability. Multi-model fallback is not paranoia at this stage — see vendor resilience and multi-model fallback strategy.
What this means for teams in MENA
Two things, specifically.
Token cost has been the quiet ceiling on agent adoption across the region. A workflow that costs $400/month in inference is an easy approval in San Francisco and a hard one in Tunis or Casablanca. Models in this price and capability band move a meaningful set of automation projects from "interesting, later" to "fundable now."
And the MIT license removes the compliance conversation that usually stalls open-weight adoption in regulated sectors. Banking, health, and public-sector clients who cannot send data to a US-hosted API now have a path that does not require a bespoke vendor agreement — provided someone has budgeted for the hardware.
The honest summary
DeepSeek-V4-Flash-0731 is a re-post-trained checkpoint with vendor-reported benchmarks, a genuinely permissive license, aggressive cache pricing, and real hardware requirements if you want to run it yourself. It is not a frontier model and does not claim to be.
What it is, is strong evidence that the useful question in 2026 is no longer "which model is smartest" but "what is the cheapest model that clears the bar for this specific step." Run the eval on your own traces. The answer for your workload will be more informative than any leaderboard.
Building agent workflows and want the cost model to hold up in production? Noqta helps teams across Tunisia and the Gulf design AI systems that ship — and stay affordable at scale.