DeepSeek V4.1 Flash: Benchmarks, API Pricing, Size and Setup

TL;DR

Last updated November 9, 2026.

DeepSeek V4.1 Flash is DeepSeek's current Flash model. It launched on September 10, 2026, and DeepSeek updated it to DeepSeek-V4.1-Flash-1017 on October 17. It is a mixture-of-experts model with a 552B backbone plus 196B of Engram conditional memory. It activates 8B parameters per token during prefill and 16B during decode. It reads text and images, has a 1M-token context, and ships under the MIT license. In DeepSeek's own table, the 1017 update scores 77.0 on DeepSWE v1.1, ahead of GLM-5.3, GPT-5.6 Sol and Claude Opus 5, and 42.3 on Terminal-Bench 4.0, behind Opus 5's 51.8. On Artificial Analysis's Intelligence Index v5 it scores 32, the same as GLM-5.3 and 14 points behind Kimi K3.1 Pro. DeepSeek's API charges $0.15 per million input tokens, $0.60 for output and $0.003 for cache hits off-peak, and double that at peak. Morph serves it as morph-dsv41flash at $0.12 per million input tokens, $0.008 cached and $0.80 output, the same at every hour. There is no V4.1 Pro yet.

What Is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is the model behind DeepSeek's deepseek-flash API name. DeepSeek released it on September 10, 2026 as "the smallest model in our new architecture family, with native visual understanding." On October 17 it shipped DeepSeek-V4.1-Flash-1017, which the model card calls "a minor update" with gains in reasoning, coding, agent tasks and vision. DeepSeek says 1017 also produces shorter outputs and finishes tasks faster. Since that date, deepseek-flash serves 1017 at unchanged prices.

It replaced V4 Flash on DeepSeek's API. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the old models are retired and those requests are served by 1017 at the Flash price. For the original V4 Flash (0731) and its numbers, see the DeepSeek V4 Flash page.

There is no DeepSeek V4.1 Pro. DeepSeek's September 10 post said deepseek-v4-pro would route to V4.1 Flash from September 14 "until V4.1-Pro launches." It then reversed that and kept serving V4 Pro. The current deepseek-v4-pro model is DeepSeek-V4-Pro-0813, and the deepseek-ai Hugging Face org has no V4.1 Pro repo. The release history, the architecture changes from V4, and the V4.1 Pro status are covered on the DeepSeek V4.1 page. This page covers V4.1 Flash as a product: what it scores against peers, what it costs, and how to call it.

DeepSeek V4.1 Flash Size: Parameters, Weights and Context

DeepSeek's release post calls V4.1 Flash a "552B-parameter MoE." The model card is more precise: 552B backbone parameters, plus a separate 196B Engram conditional memory that the model reads sparsely by token lookup. Only 8B parameters are active per token while it reads your prompt (prefill) and 16B while it writes (decode). That small active count is why it is priced like a small model.

Two numbers trip people up. First, the Hugging Face safetensors widget on the 1017 repo shows a larger parameter total. That widget counts every tensor in the files (BF16, F32, F8_E4M3 and I8 types), and it is not DeepSeek's headline figure. Use "552B backbone plus 196B Engram" when you compare sizes. Second, "384K" and "256K" are different settings: 384K is the output cap on DeepSeek's API, and 256K is the card's recommended max_tokens when you run the weights yourself.

The weights ship without a Jinja chat template. DeepSeek provides an encoding folder and the deepseek-recipe library to turn Chat Completions, Responses and Messages requests into its prompt format. How the causal encoder-decoder, CSA2 attention and Engram memory work, and what it takes to serve the open weights, are on the DeepSeek V4.1 page.

DeepSeek V4.1 Flash Benchmarks vs GLM-5.3, Kimi, GPT-5.6 Sol and Claude

Almost every V4.1 Flash benchmark number comes from DeepSeek. Its model card says the instruct results use maximum effort with temperature 1.0 and top_p 0.95, and DeepSeek picked the harness for each coding test: its own DeepSeek Harness for most September rows, Claude Code and mini-SWE-v2 for the 1017 headline rows. The peer columns below are the numbers DeepSeek published next to its own at release, so they compare against Kimi K3 and Claude Opus 5, which have since been succeeded by Kimi K3.1 Pro and Claude Opus 5.5. DeepSeek reports no SWE-bench Verified, SWE-bench Pro or LiveCodeBench score for V4.1 Flash.

DeepSeek ran Terminal-Bench 4.0 in Claude Code and DeepSWE in mini-SWE-v2 for the 1017 column. For the September checkpoint, NL2Repo-Bench is 65.4 in DeepSeek's release notes and change log but 64.0 on the Hugging Face card; we show the release notes figure.

The September launch table also covered rows that the 1017 table dropped. These apply to the September checkpoint only:

Kimi K3.1 Pro vs V4.1 Flash 1017 (Moonshot's table)

The newest head-to-head comes from Moonshot. Its Kimi K3.1 Pro model card includes a V4.1 Flash 1017 column. Moonshot ran K3.1 Pro in Claude Code and took most other scores from vendor posts or official leaderboards. Moonshot cites the AutomationBench and GDPval rows from Artificial Analysis, and ran SWE Marathon for V4.1 Flash 1017 itself in Claude Code.

Two things to notice. Moonshot lists GLM-5.3 at 41.8 on Terminal-Bench 4.0, while DeepSeek's table lists 37.9: same benchmark, different reported runs, which is why we keep the vendor tables apart. And the DeepSWE row is closer than it looks. K3.1 Pro's 75.5 was run in Claude Code. V4.1 Flash's 77.0 was run in DeepSeek's chosen harness.

The harness changes the score

DeepSeek tested the September checkpoint in eight agent scaffolds with the same settings. On DeepSWE v1.1 it scored 74.2 in mini-SWE, 69.8 in Claude Code and 65.6 in Codex. On Terminal-Bench 2.1 it ranged from 84.1 (Codex) to 90.6 (DeepSeek Harness Minimal). If you run V4.1 Flash in Claude Code or Codex, expect results nearer the lower numbers. The full scaffold table is on the DeepSeek V4.1 page.

Independent numbers

Little has been measured outside DeepSeek, and the main independent score puts V4.1 Flash well below the frontier. Artificial Analysis's Intelligence Index v5, published October 27, 2026, scores V4.1 Flash 1017 (max) at 32. Among models with released weights on that date, it tied GLM-5.3 (max) for the top spot, 26 points behind the overall leader, Claude Opus 5.5 (max) at 58. It also sits below Claude Haiku 5.5 (max) at 36. Kimi K3.1 Pro, whose weights were not yet out, scored 46.

On speed, Artificial Analysis lists V4.1 Flash (Max) at 231 output tokens per second and 47.05 seconds to the first answer token, and the non-reasoning variant at 226 tokens per second and 1.30 seconds (median across providers), as of November 9, 2026.

Cheap Tokens, Expensive Habits: Effort Levels and Cache Hits

The per-token price is half of what you pay. The other half is how many tokens the model spends. Two settings decide that: the reasoning effort you ask for, and how much of each prompt hits the cache.

Effort

DeepSeek's headline scores use maximum effort. Its API defaults to high effort with thinking on, and accepts low, high and max. Max is slow: Artificial Analysis measured 47.05 seconds to the first answer token for V4.1 Flash at Max, against 1.30 seconds for the non-reasoning variant. That gap is the thinking pass, and you pay for it in time and tokens. DeepSeek says the 1017 update writes shorter outputs than the September checkpoint. Our sources include no independent output-token count for 1017, so that claim is untested. Use the lowest effort that passes your evals; the thinking-mode rules are on the DeepSeek V4.1 page.

Cache hits

DeepSeek pitches V4.1 Flash on cache costs. Its release post says "cache-hit charges often account for a large share of agent costs," and V4.1 Flash's KV cache needs 1/4 the HBM and 1/8 the SSD storage of the previous generation. On DeepSeek's API, a cache hit costs 2% of a miss. On Morph it costs 6.7% ($0.008 vs $0.12), and Morph's cached rate is 2.7x DeepSeek's off-peak cached rate. So the cheaper provider depends on your traffic shape:

On the cached agent turn, cache hits are 26% of the Morph bill and 13% of the DeepSeek off-peak bill. Fresh input is where Morph's $0.12 rate pays off. Long outputs favor DeepSeek's off-peak $0.60.

Timing matters too. DeepSeek's peak window is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, 35 of the week's 168 hours (20.8%). In Beijing time that is 09:00 to 12:00 and 14:00 to 18:00. A 9 to 5 workday in US Pacific or Eastern time falls entirely in DeepSeek's off-peak window. Morph charges one rate at every hour. For how cache hits are counted and why switching models mid-session resets them, see prompt caching.

DeepSeek V4.1 Flash Pricing Across Providers

V4.1 Flash is MIT-licensed, so many hosts serve it. On November 9, 2026, OpenRouter listed 50 endpoints for deepseek/deepseek-v4.1-flash-1017 and 37 for the September slug deepseek/deepseek-v4.1-flash. DeepSeek itself appears only on the 1017 slug, because it now serves only 1017. Here is the spread, per 1M tokens:

Read the table three ways. Input runs from $0.006 to $0.60 across the 50 endpoints, and output from $0.15 to $2.40, so "the V4.1 Flash price" depends on who you call. Nine of the 50 endpoints list $0.30 input and $1.20 output, which is DeepSeek's peak rate. And OpenRouter shows each host's declared quantization (fp4, fp8, mxfp4 or none declared), so check that column before you pick the cheapest row.

Morph is not the cheapest V4.1 Flash provider. Against DeepSeek's own API, Morph's $0.12 input is lower than both DeepSeek rates, its $0.80 output sits between DeepSeek's off-peak and peak rates, and its $0.008 cache hit is higher than both. Morph is also one of the OpenRouter providers for both slugs; prices on this page are Morph's direct API. What Morph adds is covered in running it on Morph.

DeepSeek V4.1 Flash API: How to Call It

Both Morph and DeepSeek expose V4.1 Flash through an OpenAI-compatible API, so calling it is a base URL and a model string. On Morph the base URL is https://api.morphllm.com/v1 and the model is morph-dsv41flash. Create a key on the API keys page and send it as a Bearer token.

OpenAI SDK (Python)

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key=os.environ["MORPH_API_KEY"],
)

resp = client.chat.completions.create(
    model="morph-dsv41flash",
    messages=[
        {"role": "user", "content": "Find the bug in this function and explain it in two sentences: ..."},
    ],
    extra_body={"reasoning_effort": "low"},  # cap the thinking pass on cheap tasks
)
print(resp.choices[0].message.content)
print(resp.usage)

The V4.1 Flash launch post lists the reasoning_effort values Morph accepts and shows an image request.

curl

curl https://api.morphllm.com/v1/chat/completions \
  -H "Authorization: Bearer $MORPH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "morph-dsv41flash",
    "messages": [{"role": "user", "content": "Summarize this diff in three bullets."}]
  }'

DeepSeek's own API

To call DeepSeek directly, use base URL https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format) and model deepseek-flash. Thinking is on by default. Turn it off with {"thinking": {"type": "disabled"}}, or set reasoning_effort to low, high or max. One rule breaks agents that ignore it: when a request carries tools, pass back the reasoning_content of earlier turns.

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key=os.environ["DEEPSEEK_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Classify this ticket: ..."}],
    extra_body={"thinking": {"type": "disabled"}},
)

For more on DeepSeek's API (keys, cache pricing, older models), see the DeepSeek API guide.

Use DeepSeek V4.1 Flash in Claude Code and Codex

Claude Code

Morph serves the Anthropic Messages API at https://api.morphllm.com, so Claude Code needs four environment variables to run on V4.1 Flash:

export ANTHROPIC_BASE_URL="https://api.morphllm.com"
export ANTHROPIC_AUTH_TOKEN="YOUR_MORPH_API_KEY"
export ANTHROPIC_MODEL="morph-dsv41flash"
export ANTHROPIC_SMALL_FAST_MODEL="morph-dsv41flash"
claude

Morph's Ultra setup pairs Kimi K3.1 for the main loop with V4.1 Flash for Claude Code's background calls, such as titles and summaries:

export ANTHROPIC_BASE_URL="https://api.morphllm.com"
export ANTHROPIC_AUTH_TOKEN="YOUR_MORPH_API_KEY"
export ANTHROPIC_MODEL="morph-kimik31"
export ANTHROPIC_SMALL_FAST_MODEL="morph-dsv41flash"
claude

Codex

Codex talks to providers over the Responses API. In ~/.codex/config.toml, set the two top-level keys above the first table, then add the provider:

model = "morph-dsv41flash"
model_provider = "morph"

[model_providers.morph]
name = "Morph"
base_url = "https://api.morphllm.com/v1"
env_key = "MORPH_API_KEY"
wire_api = "responses"

If your default is morph-kimik31, run codex -m morph-dsv41flash when you want the faster model for one session. Keep the harness numbers in mind: on DeepSWE v1.1, DeepSeek measured the September checkpoint at 65.6 in Codex, 8.6 points below its mini-SWE score and one of the two lowest of the eight scaffolds it tested.

One-command agent setup

Morph's agent setup command:

curl -fsSL https://www.morphllm.com/install.sh | bash && morph install

Running DeepSeek V4.1 Flash on Morph: Speed, Rates and the Ultra Plan

Morph serves V4.1 Flash as morph-dsv41flash at $0.12 per million input tokens, $0.008 cached and $0.80 output, the same at every hour, with a 1M context. morph-dsv41flash-1017 routes to the same deployment at the same price.

Speed. Morph's public hourly charts at ultra.morphllm.com show about 215 output tokens per second and about 0.6 seconds to first token for V4.1 Flash (median of hourly medians, 24 hours ending November 9, 2026). Artificial Analysis lists 231 tokens per second for V4.1 Flash (Max), so Morph is not the fastest published number for this model. The two sources measure differently, so compare them loosely.

One key, three open models. The same Morph key calls morph-dsv41flash, morph-kimik31 (Kimi K3.1 Pro) and morph-glm53-744b (GLM-5.3) through the OpenAI-compatible API at https://api.morphllm.com/v1 or the Anthropic Messages API at https://api.morphllm.com. Morph's 50% standby tier applies to Kimi K3.1 and GLM-5.3, not to V4.1 Flash.

Ultra plan. Ultra costs $200 a month and includes unlimited Kimi K3.1 and V4.1 Flash with no token meter. Morph caps it at 1,000 subscriptions. Each subscription can use 2,103,237 uncached input tokens and 76,559 output tokens per minute on V4.1 Flash. At metered rates, 1,000 of the example agent turns above cost $2.96, so Ultra pays for itself on V4.1 Flash alone after about 67,600 such turns a month, about 2,250 a day. Kimi K3.1 use on the same plan brings that point much closer. At that turn shape, the per-minute limits allow about 38 turns a minute.

DeepSeek V4.1 Flash vs Kimi K3.1, GLM-5.3 and Claude

All seven models below have a 1M-token context. They differ by an order of magnitude in price and by license. DeepSWE v1.1 is the one coding benchmark that DeepSeek, Moonshot, Z.ai and Anthropic all publish, but each vendor ran it its own way, so treat that column as a rough guide.

Price is where V4.1 Flash separates. Same request shape, every model at list price:

Real runs will not use identical token counts: tokenizers and verbosity differ by model. Moonshot and Anthropic also bill cache writes, which this table leaves out. Moonshot's first-party K3.1 price is below Morph's; Morph's K3.1 case is speed, the Ultra plan and standby pricing (see Kimi K3.1 API).

Quality is less one-sided. On Terminal-Bench 4.0, V4.1 Flash 1017 trails Kimi K3.1 Pro and Opus 5.5 by more than 20 points in Moonshot's table. On DeepSWE v1.1 its 77.0 is the highest vendor-reported number in the table above, with the harness caveats above. On Artificial Analysis's Intelligence Index v5 it scores 32, level with GLM-5.3 and 14 points behind Kimi K3.1 Pro.

The closed model to test it against is Claude Haiku 5.5. Haiku 5.5 lists $0.10 input and $0.50 output for prompts up to 100,000 tokens, below DeepSeek's off-peak rates, and it scores 36 on Intelligence Index v5 against V4.1 Flash 1017's 32. Past 100,000 tokens Haiku 5.5 rises to $0.50 input, $0.05 cached and $2.50 output, and V4.1 Flash costs less on every rate. On these numbers, Haiku 5.5 wins short prompts on both price and score, and V4.1 Flash wins long-context work on price.

Deep dives: Kimi K3.1, GLM-5.3, GLM-5.3 vs Kimi K3.1 and Kimi K3.1 vs Claude.

Pros and Cons

When to Use DeepSeek V4.1 Flash (and When Not)

For the 1017 update's model IDs and billing on Morph, see the V4.1 Flash 1017 page. For what changed from V4 and how to serve the weights, see DeepSeek V4.1.

Frequently Asked Questions