DeepSeek V4.1 Flash: Current API Pricing, Architecture and Local Limits

Checked October 9, 2026. DeepSeek V4.1 Flash has been available as deepseek-flash in the DeepSeek API since September 10. It combines a one-million-token context window, native vision and an asymmetric Causal Encoder-Decoder architecture built around a 552B MoE backbone. ¹²

Technical specifications

Item	DeepSeek V4.1 Flash
API model	deepseek-flash
Context	up to 1M tokens
Maximum API output	up to 384K tokens in DeepSeek’s current first-party rate card; providers and endpoints can impose narrower limits
Architecture	Causal Encoder-Decoder, MoE
Backbone	552B parameters
Active on input	about 8B parameters per token
Active on output	about 16B parameters per token
Vision	yes
Tool calling	yes
Responses API	yes
Open weights	yes

DeepSeek also describes a large Engram memory component. The published Hugging Face files are much larger than the 8B/16B activation figures suggest. Those numbers measure different layers of the system and should not be collapsed into the claim that this is a “16B model.” ³⁴

Why the KV cache matters for agents

DeepSeek says V4.1 Flash reduces the global KV-cache footprint to one quarter of the HBM and one eighth of the SSD storage used by V4 Flash. The model card also describes roughly 890 bytes of global KV cache per token. ¹³

For agents that repeatedly work over repositories or long document sets, that can matter more than the headline context maximum. A one-million-token window is still expensive when every request introduces a large amount of new text. Caching helps most when the input is genuinely reused.

API pricing

DeepSeek continues to use peak and off-peak pricing, with off-peak rates at half the peak rate. Current list prices are: ⁵

Token class	Off-peak	Peak
Cache hit	$0.003 / 1M	$0.006 / 1M
Uncached input	$0.15 / 1M	$0.30 / 1M
Output	$0.60 / 1M	$1.20 / 1M

DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Weekends and Chinese public holidays are fully off-peak. ⁵

Agent costs with and without prompt-cache hits

DeepSeek’s first-party API rate card bills uncached input, verified cached input and output at different rates. The same input token must not be billed twice as both cache hit and miss. Here is a fixed-token illustration: 10M input tokens total, of which 8M are genuine cache hits, plus 2M output tokens.

Our calculated example	Off-peak	Peak
No hits: 10M uncached + 2M output	$2.70	$5.40
Cached: 2M uncached + 8M cached + 2M output	$1.524	$3.048

Cached-case math: Off-peak 2 × $0.15 + 8 × $0.003 + 2 × $0.60 = $1.524; peak 2 × $0.30 + 8 × $0.006 + 2 × $1.20 = $3.048. These are rate-card calculations, not observed Mac speed tests or measured agent bills. Check the API’s actual cached-token accounting. Retries, external tools, provider routing, taxes and other usage can change the result.

Peak windows and time zones: DeepSeek defines peak on weekdays as 01:00–04:00 UTC and 06:00–10:00 UTC, excluding Chinese public holidays. All weekends and Chinese public holidays are off-peak. In the U.S., local-clock peak windows vary with time zone and daylight saving; schedule automation against UTC rather than assuming one permanent U.S. clock time. These are first-party DeepSeek API rates, not guaranteed third-party router prices.⁵

Cheap tokens do not guarantee cheap completed tasks. Reasoning length, retries and tool loops can erase the list-price advantage. Agent evaluations should therefore track cost per successfully completed task.

Reproduce the cost of a cache-heavy coding agent

These calculations use DeepSeek’s direct API prices only, not OpenRouter or a Harness subscription. Count a token as cached only when DeepSeek actually bills it as a cache hit. This Python script uses the three distinct token prices on the official rate card. It does not need an API key, make network calls, or consume credits.

RATE=offpeak UNCACHED=2000000 CACHED=8000000 OUTPUT=2000000 python3 - <<'PY'
from decimal import Decimal
import os

# Only a rate-card calculator; never accesses the DeepSeek API.
rate = os.environ.get("RATE", "offpeak")
if rate not in ("offpeak", "peak"):
    raise SystemExit("RATE must be offpeak or peak")
prices = {
  "UNCACHED": Decimal("0.30" if rate == "peak" else "0.15"),
  "CACHED": Decimal("0.006" if rate == "peak" else "0.003"),
  "OUTPUT": Decimal("1.20" if rate == "peak" else "0.60"),
}
defaults = {"UNCACHED": "2000000", "CACHED": "8000000", "OUTPUT": "2000000"}
costs = {}
for kind, price in prices.items():
    count = Decimal(os.environ.get(kind, defaults[kind]))
    if not count.is_finite() or count < 0 or count != count.to_integral_value():
        raise SystemExit("Invalid token count for " + kind)
    costs[kind] = count * price / Decimal(1000000)
    print(kind, "USD", f"{costs[kind]:.4f}")
print("TOTAL USD", f"{sum(costs.values(), Decimal(0)):.4f}")
print("Illustrative only; verify actual cache hits and vendor billing.")
PY

The defaults cost $1.5240 off-peak, or $3.0480 using RATE=peak. With no cache hits (UNCACHED=10000000 CACHED=0), the same workload instead costs $2.7000 off-peak or $5.4000 peak. These are rate-card calculations, not measured agent workloads. Include real retries, tool calls, prompts and invoice token counts when evaluating cost per completed task.

Vision, tools and the Responses API

V4.1 Flash includes native vision. DeepSeek’s vision guide lists JPEG, PNG, GIF and WebP, with image input supplied by URL, base64 or the Files API. Tool calls and the Responses API are also documented. ⁶⁷⁸

That combination makes deepseek-flash useful for coding and research agents that need to inspect screenshots, document pages or UI state while also calling tools.

Vision limits: do not use the retired Vision Exp cap

The discontinued V4 Flash Vision Exp used a 384-image-token ceiling. For today’s deepseek-flash, DeepSeek documents up to 1,024 tokens per image after resizing. That is not a flat cost for every image. These are limits from DeepSeek’s direct Vision API, not necessarily what third-party providers expose.

Input path	Published limit	Integration caveat
External URL / base64	32 MiB per image	Inline base64 counts toward the 48 MiB request-body limit
Files API file_id	64 MiB per image	Reuse uploaded files rather than sending bytes each time
Per request	600 images	Up to 64 MiB without file_id, or 200 MiB including files
Image dimensions	8,192 pixels per side	Reduced to 4,096 pixels per side for 15+ images
Chat Completions	Image in user only	system or assistant images yield HTTP 400
Responses API	input_image	Provide either image_url or file_id

The API accepts JPEG, PNG, GIF and WebP, determined from actual image contents. detail=low downsamples to 512 × 512 pixels; high and original preserve resolution. detail is ignored for file_id. A file on your Mac is not a publicly reachable URL: send it as base64, provide an accessible HTTPS URL, or upload it through the Files API first. No private images were uploaded or paid requests made for this guide.

Important correction: V4 Pro is still available

The September 10 launch announcement initially said deepseek-v4-pro would be routed to V4.1 Flash from September 14. That plan changed.

DeepSeek’s official changelog now says V4 Pro will continue to be offered as an API service due to user demand, with billing unchanged, and that any future change will be announced separately. ⁹

This is why launch posts should not be treated as permanent status pages for fast-moving model catalogs.

Benchmarks

DeepSeek reports 90.6 on Terminal-Bench 2.1, 90.9 on GPQA Diamond and 74.2 on DeepSWE v1.1 among its launch evaluations. These are first-party numbers. ⁹

Independent measurements should remain separate because prompting, harness design, reasoning settings and provider latency can materially change results. ¹⁰

Local use on a Mac

Open weights do not make the complete checkpoint small. The Hugging Face repository is in the hundreds of gigabytes. Even a theoretical four-bit lower bound for the published safetensor parameter count sits far above the memory capacity of ordinary Macs. ³⁴

The 8B and 16B active-parameter figures describe how much of the MoE participates per token. They do not mean only 8B or 16B weights need to reside in memory. For a normal Mac, V4.1 Flash is primarily an API model.

Connect DeepSeek V4.1 Flash to DeepSeek Harness from a Mac

The model and the agent runtime are different products. V4.1 Flash generates responses via the hosted DeepSeek API. DeepSeek Harness (DSH) supplies workspace access, sessions, tools and MCP around a configured LLM. Using Harness does not mean that the 552B MoE weights now execute locally on Apple Silicon, nor does it bypass an API quota. See our companion DeepSeek Harness setup and safety guide for the complete runtime walkthrough.¹¹

Five checks before trusting a coding-agent workflow

1. Launch the Web UI using DeepSeek’s official setup guide and select a disposable workspace. Do not grant access to your whole home directory or production SSH keys.
2. Under Settings → Models, set up the built-in DeepSeek provider with an actual DeepSeek API key. A consumer app subscription is not an API credential and an API key does not install local model weights.
3. Verify the exact available model ID in the picker and your developer account. DeepSeek uses the deepseek-flash API alias for its Flash route; check which model version that alias currently resolves to before publishing a benchmark or choosing V4 Pro. A provider’s generic model name is not proof of a pinned checkpoint.⁹¹¹
4. Start a new Session with a safe synthetic code-review task. Confirm a text response before allowing read-only tools on a throwaway repo. Model-list discovery alone does not verify file permissions, multimodal input, reasoning settings or successful tool execution.
5. Compare cost per successfully completed task, reported prompt/output tokens, retry/tool failures and total latency. Reconcile charges against the real DeepSeek dashboard; the cache-read price applies only to input tokens actually billed as cached. Keep API keys out of repos, logs and screenshots.

Scheduled work is different from reminders. DSH Host Schedule stores session-linked reminders; delivery is not a receipt for a successful CI run. For a separate scheduler such as Hermes to run unattended coding jobs, it must explicitly submit an ACP worker prompt, validate tool outcomes and only mark the task done after checks pass. This is a proposed controller architecture, not a tested or built-in Hermes↔DSH integration.¹²

Privacy limit: Running a DSH browser UI on your Mac does not make DeepSeek cloud inference local, U.S.-hosted or compliant with a particular privacy law. Check your own processing terms and retention requirements before uploading regulated or client content. No paid live API calls or model benchmarks were performed for this procedure.

Current role of V4.1 Flash

DeepSeek V4.1 Flash combines very low token prices, one-million-token context, native vision and an architecture explicitly designed for cache-efficient agent work. The API picture is also clearer than it was on launch day: V4 Flash and V4 Flash Vision Exp are retired and routed to V4.1 Flash, while V4 Pro remains a separate service after DeepSeek reversed its original routing plan. ¹⁹⁵

For Mac users, the split is straightforward: excellent as a cloud model to evaluate, unrealistic as a full local checkpoint except on unusually large inference systems.

Source references

Footnotes

1. DeepSeek: launch ↩ ↩² ↩³

2. DeepSeek: API update ↩

3. DeepSeek: model card ↩ ↩² ↩³

4. DeepSeek: weights ↩ ↩²

5. DeepSeek: API pricing ↩ ↩² ↩³ ↩⁴

6. DeepSeek: vision API ↩

7. DeepSeek: Responses API ↩

8. DeepSeek: tool calling ↩

9. DeepSeek: changelog ↩ ↩² ↩³ ↩⁴

10. Artificial Analysis ↩

11. DeepSeek Harness provider guide ↩ ↩²

12. DeepSeek Harness scheduling guide ↩