DeepSeek V4: Facts, Claims, and Fit

Your evaluation log says deepseek-v4-pro. The architecture report says 1.6 trillion parameters. Hugging Face displays 1.7 trillion. An older run used the same API name for a different release. None of those facts necessarily contradicts the others. They describe different parts of the system you tested.

DeepSeek V4 is a useful case study in reading a model release without turning its strongest number into a deployment recommendation. The checkpoints, long-context architecture, and hosted prices are concrete. Whether they improve your coding assistant depends on what survives a comparison with the same tasks, tools, and budgets.

Scope and date: V4 Flash 0731 and Pro 0813 checkpoints, with hosted-service changes checked on October 10, 2026. Since September 10, deepseek-v4-flash serves V4.1 Flash. The architecture and benchmark sections below describe the older V4 checkpoints; the billing example uses today's hosted models.^{[1]Reference 1DeepSeek API Change Loghttps://api-docs.deepseek.com/updates/}

One API name, several releases

V4 Preview arrived on April 24. Flash received new post-training on July 31, while Pro reached general availability on August 13. The hosted names stayed the same through those upgrades.^{[2]Reference 2DeepSeek V4 Preview Releasehttps://api-docs.deepseek.com/news/news260424[1]Reference 1DeepSeek API Change Loghttps://api-docs.deepseek.com/updates/[3]Reference 3DeepSeek-V4-Pro GA Releasehttps://api-docs.deepseek.com/news/news260813/}

The current pricing page maps deepseek-flash to V4.1 Flash and deepseek-v4-pro to Pro 0813. It accepts deepseek-v4-flash and deepseek-v4-flash-vision-exp as compatibility names for V4.1 Flash, billed at the new Flash rate.^{[4]Reference 4Models and Pricinghttps://api-docs.deepseek.com/quick_start/pricing} Under this mapping, a run with the legacy Flash string measures V4.1 Flash. Use dated results or response metadata to identify any earlier Flash 0731 run. Record the documented mapping and evaluation date.

DeepSeek initially planned to end V4 Pro API service on September 14, but its September 10 changelog says Pro will continue with unchanged billing until further notice.^{[1]Reference 1DeepSeek API Change Loghttps://api-docs.deepseek.com/updates/} Check the service lifecycle before planning a migration around an older announcement.

For an open checkpoint, record the repository revision, file hashes, tokenizer/encoding implementation, and serving configuration. You can freeze those artifacts, unlike a mutable hosted alias, but they still don't reproduce the hosted service's entire runtime. Freezing the checkpoint doesn't settle the parameter count. Why does the report say 1.6 trillion parameters when the Hub lists 1.7 trillion?

Why the parameter counts differ

The report describes two mixture-of-experts models: Flash has 284B total parameters with 13B active per token; Pro has 1.6T total with 49B active.^{[5]Reference 5DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligencehttps://arxiv.org/abs/2606.19348} An MoE activates only some expert networks for each token. Active parameters help explain compute, while the full weight pool still requires storage and a residency or offloading plan. Neither count alone predicts latency. Mixture of Experts Architecture explains the routing and communication costs.

The newer checkpoint packages also include DSpark, a speculative-decoding module. It proposes tokens for the target model to verify. DeepSeek's DSpark card describes an attached draft path, and the dated Flash and Pro cards identify that module in their packages.^{[6]Reference 6DeepSeek-V4-Flash-DSpark Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark[7]Reference 7DeepSeek-V4-Flash-0731 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731[8]Reference 8DeepSeek-V4-Pro-0813 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813}

The July changelog explicitly says Flash 0731 retains the preview architecture and size. The Pro 0813 card likewise says it retains the preview Pro structure with DSpark attached.^{[1]Reference 1DeepSeek API Change Loghttps://api-docs.deepseek.com/updates/[8]Reference 8DeepSeek-V4-Pro-0813 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813} The Hub badges count a different artifact than the target-architecture figures. Don't subtract rounded badges to infer an exact draft-module size, or interpret the larger badge as evidence that the target became wider.

This distinction also affects measurement. Speculation adds draft work and verification overhead. Its speedup depends on acceptance and the serving workload. Speculative Decoding explains how accepted drafts can make a larger packaged checkpoint generate faster. Measure speed on the target serving workload. How does a million-token window avoid storing every past key and value?

What makes the long-context design different

The preview models support a reported one-million-token context window. Their attention layers combine two ways to reduce the number of key-value cache entries used for long-range context:

• Compressed Sparse Attention (CSA): compress the sequence, then select a subset of compressed entries for each query.
• Heavily Compressed Attention (HCA): compress more aggressively, then attend densely over that smaller representation.

The report uses a fourfold sequence compression rate for CSA and 128-fold for HCA, with an additional sliding-window path for recent local detail. CSA and HCA layer types alternate through the stack, alongside the local-detail path.^{[5]Reference 5DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligencehttps://arxiv.org/abs/2606.19348} These compression rates describe V4. Verify V4.1 separately even when a legacy API name still works.

The diagram separates the two layer types. Each has its own compressed long-range path and a recent-token path; later layers alternate the long-range mechanism.

View larger

DeepSeek's estimated costs at 1M context are plotted below. FLOPs are single-token inference work expressed in equivalent FP8 operations; KV cache is accumulated cache size. Measure request latency with a serving benchmark.^{[5]Reference 5DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligencehttps://arxiv.org/abs/2606.19348}

View larger

The baseline is DeepSeek-V3.2, which already uses sparse attention. The reductions combine compression, sparsity, precision choices, and different model configurations, so interpret them as joint architecture estimates.

A small cache says little about whether a model can retrieve a buried identifier or reconcile two distant policy clauses. Test those cases at your actual input lengths. Measure prefill, decode, memory, and end-to-end latency separately. KV Cache and PagedAttention separates cache capacity from allocation and scheduling. Evaluate model quality using the model card's harness, sampling settings, and date before making a choice.

Read benchmark rows with their conditions attached

The Pro 0813 card reports these selected results. DeepSeek published the measurements; we haven't independently reproduced them.^{[8]Reference 8DeepSeek-V4-Pro-0813 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813}

The public code-agent evaluation notes specify DeepSeek Harness minimal mode, max reasoning effort, temperature=1.0, and top_p=0.95. The cards also include two internal DSBench datasets, which outside readers can't fully inspect.^{[7]Reference 7DeepSeek-V4-Flash-0731 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731[8]Reference 8DeepSeek-V4-Pro-0813 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813}

One subtle reproducibility problem: DeepSeek's hosted thinking-mode API ignores caller-supplied temperature. Its current guide says top_p takes effect in thinking mode within 0.95 to 1.0, with lower values clamped to 0.95. Non-thinking mode fixes it at 1.0.^{[9]Reference 9Thinking Modehttps://api-docs.deepseek.com/guides/thinking_mode} Reproduce a benchmark by pinning its checkpoint, harness, and sampling implementation. Record the endpoint's effective settings alongside the JSON your client sends.

Under the vendor's reported setup, Pro 0813 scores above Flash 0731 on these four rows. Comparing cost per successful task, latency, or rank across providers requires comparable tasks, harnesses, retry limits, and acceptance criteria. These dated checkpoint results provide none of those measurements. Consult current billing documentation for hosted model prices and route names.

Turn the rate card into a workload bill

Before you treat a higher score as a reason to pick a lane, price the hosted models those names point to now. As of October 10, deepseek-flash (V4.1 Flash) and deepseek-v4-pro (Pro 0813) advertise 1M context, maximum output of 384K tokens, thinking and non-thinking modes, tool calls, JSON output, Responses API support, and an Anthropic-compatible interface. Flash supports vision. Pro doesn't.^{[4]Reference 4Models and Pricinghttps://api-docs.deepseek.com/quick_start/pricing}

Those limits share one budget. The Chat Completions reference limits input plus generated tokens to the model's context length. Its 393,216-token output ceiling shares that context window. Reserve room for reasoning and the final answer before packing documents. A 1M-token input cannot also request 384K generated tokens.^{[10]Reference 10Create Chat Completionhttps://api-docs.deepseek.com/api/create-chat-completion}

Peak/off-peak pricing took effect at 16:00 UTC on August 16.^{[3]Reference 3DeepSeek-V4-Pro GA Releasehttps://api-docs.deepseek.com/news/news260813/} Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour is off-peak. The rates below are USD per million tokens, checked October 10.^{[4]Reference 4Models and Pricinghttps://api-docs.deepseek.com/quick_start/pricing}

For a batch of requests totaling 2M input tokens and 0.5M output tokens, cold-cache off-peak charges are $0.60 for V4.1 Flash and $2.31 for Pro. It assumes every token uses the indicated billing period and excludes tools, retries not already counted, and human review. These prices compare today's hosted lanes; the benchmark table covers two V4 checkpoints.

The calculator uses token counts and avoids guessing a request-level cache-hit rate. A 70% input-token hit share means 1.4M cached input tokens and 0.6M uncached ones. Split a mixed peak/off-peak workload into separate usage buckets before summing. The September 23 fixture below still matches the October 10 rate card.

from dataclasses import dataclass
from decimal import Decimal

@dataclass(frozen=True)
class HostedModelRate:
    name: str
    cache_hit_per_m: Decimal
    cache_miss_per_m: Decimal
    output_per_m: Decimal

    def workload_cost(
        self,
        cache_hit_tokens: int,
        cache_miss_tokens: int,
        output_tokens: int,
    ) -> Decimal:
        counts = (cache_hit_tokens, cache_miss_tokens, output_tokens)
        if any(type(count) is not int or count < 0 for count in counts):
            raise ValueError("token counts must be nonnegative integers")
        return (
            cache_hit_tokens * self.cache_hit_per_m
            + cache_miss_tokens * self.cache_miss_per_m
            + output_tokens * self.output_per_m
        ) / Decimal(1_000_000)

# Strings preserve the published decimal rates exactly.
rates = [
    HostedModelRate("v4.1-flash-offpeak", Decimal("0.003"), Decimal("0.15"), Decimal("0.60")),
    HostedModelRate("v4.1-flash-peak", Decimal("0.006"), Decimal("0.30"), Decimal("1.20")),
    HostedModelRate("pro-offpeak", Decimal("0.022"), Decimal("0.66"), Decimal("1.98")),
    HostedModelRate("pro-peak", Decimal("0.044"), Decimal("1.32"), Decimal("3.96")),
]
for rate in rates:
    cold = rate.workload_cost(0, 2_000_000, 500_000)
    warm = rate.workload_cost(1_400_000, 600_000, 500_000)
    print(f"{rate.name}: cold=${cold:.4f}, warm(70% input hit)=${warm:.4f}")

v4.1-flash-offpeak: cold=$0.6000, warm(70% input hit)=$0.3942
v4.1-flash-peak: cold=$1.2000, warm(70% input hit)=$0.7884
pro-offpeak: cold=$2.3100, warm(70% input hit)=$1.4168
pro-peak: cold=$4.6200, warm(70% input hit)=$2.8336

These calculations apply the stated token buckets. Check provider usage after each run to observe invoice rounding and cache behavior. A lower token bill can still accompany more failed tasks or more review work. The bill also assumes the endpoint ran the settings you sent. Thinking mode, tool state, and concurrency limits can change the call while the rate card stays the same.

Compatibility doesn't remove operational differences

The current docs enable thinking by default at high effort. minimal and low map to low; medium, high, and xhigh map to high; max and ultra map to max.^{[9]Reference 9Thinking Modehttps://api-docs.deepseek.com/guides/thinking_mode} Include the effective effort in your experiment record and latency budget.

For Chat Completions requests carrying tools, preserve the assistant's reasoning_content across subsequent turns, including turns without a tool call. The guide says omitting it causes HTTP 400. Without tools, prior reasoning content is ignored.^{[9]Reference 9Thinking Modehttps://api-docs.deepseek.com/guides/thinking_mode} A compatible request format doesn't eliminate this conversation-state requirement.

Account-level concurrency limits are 2,500 for Flash and 500 for Pro, with expansion available by request. A request counts until its response completes, and excess concurrency returns HTTP 429. Multiple API keys under one account don't create independent quotas.^{[11]Reference 11Rate Limithttps://api-docs.deepseek.com/quick_start/rate_limit/} These limits bound in-flight requests. They leave token throughput unspecified, and a long reasoning request occupies a slot much longer than a short classification.

DeepSeek also documents user_id for content-safety, KV-cache, and scheduling isolation. Regular accounts still pool all user IDs under the account quota. Expanded accounts additionally enforce a per-user limit.^{[11]Reference 11Rate Limithttps://api-docs.deepseek.com/quick_start/rate_limit/} Derive that opaque ID from the authenticated tenant on your server. Letting a client choose another tenant's ID would defeat your intended isolation, and the provider field doesn't replace your application's authorization checks.

The April release note scheduled deepseek-chat and deepseek-reasoner for retirement after July 24 at 15:59 UTC.^{[2]Reference 2DeepSeek V4 Preview Releasehttps://api-docs.deepseek.com/news/news260424} Use deepseek-flash or deepseek-v4-pro for new hosted configuration. We haven't made authenticated calls to test the retired aliases or measured live availability here.

Self-hosting changes different responsibilities. Both dated checkpoint cards specify MIT licensing and provide vLLM/SGLang recipes that load the target and DSpark from the same checkpoint.^{[7]Reference 7DeepSeek-V4-Flash-0731 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731[8]Reference 8DeepSeek-V4-Pro-0813 Model Cardhttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813} They also use custom message encoding. A serving migration must validate that encoding and output parsing alongside checkpoint loading.

Record the checkpoint revision, runtime version, quantization, DSpark setting, GPU topology, and tested context length. Measure hosted equivalence and cluster capacity separately. Decide which data a provider may receive before sending evaluation traffic. Self-hosting still needs its own access, logging, and retention controls. Use a small pilot with acceptance evidence to learn which of your tasks the model finishes.

A pilot that can change your decision

Use tasks whose success you can inspect. For a repository assistant, a small starting pack might contain these cases:

Use this illustrative starting pack to develop a representative benchmark from real failures. Compare Flash, Pro, and the incumbent with equivalent permissions and acceptance tests, recording every attempt's tokens, tool costs, latency, and outcome. Give each model its supported prompt/encoding contract while keeping the task and resource budget comparable.

Keep enough identity in each result to notice a service change. For example, these two runs have the same requested model string but belong in separate evaluation groups:

Use the provider's dated changelog to identify the release mapping; a requested string alone cannot identify a hosted version.^{[1]Reference 1DeepSeek API Change Loghttps://api-docs.deepseek.com/updates/} Keep outcomes separate because the hosted release can change during the experiment. If the response omits an immutable backend version, record that limit. For self-hosted runs, add checkpoint hashes and serving configuration when comparing with hosted behavior.

If Flash succeeds on source lookup but fails the parser case, you've learned something more useful than "cheap models handle easy tasks." You have a candidate task-specific routing rule and a failure set for the next experiment. Confirm it on held-out cases before a bounded canary, and keep a rollback route that respects the same data policy.

LLM Cost Engineering develops the usage ledger. Model Gateways, Routing, and Fallbacks covers enforcing those routing contracts.

DeepSeek's releases give teams specific alternatives to measure. They don't establish that US labs are being displaced, or that one provider is the right destination for every high-risk task. A deployment decision needs the dated artifact, effective runtime settings, and accepted work per dollar. The launch narrative can't supply those last measurements.