DeepSeek V4.1 Flash: Specs, Pricing, V4-Pro Status & Local Setup

DeepSeek released V4.1-Flash on 10 September 2026 with three claims: it beats the company's own flagship, V4-Pro, while costing a fraction as much; the old Flash models are retired; and V4-Pro would be phased out. The first two held up. The third changed within days. DeepSeek said it would send all deepseek-v4-pro traffic to V4.1-Flash from 14 September, then cancelled that plan after users pushed back. If you read that V4-Pro is deprecated, the official change log says otherwise.

This guide covers the specs, the exact API model name and prices, what the V4-Pro situation means for existing integrations, which benchmark numbers hold up under independent testing (including the eye-catching 90.6 on Terminal-Bench 2.1), and what running it locally requires. That last part is more than the "Flash" name suggests. All figures were checked against DeepSeek's docs, the Hugging Face model card and third-party leaderboards on 22 September 2026.

What is DeepSeek-V4.1-Flash?

V4.1-Flash is the first and smallest model in a new DeepSeek architecture family. It is not a fine-tune of V4-Flash. The model card describes a Causal Encoder-Decoder (CED) design: a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder. The decoder builds its global KV cache from the encoder's final hidden states, so the model activates only 8B parameters per token while reading input (prefill) and 16B while generating (decode). Agent workloads are mostly input (long tool outputs, file contents, repeated context), so a cheaper prefill cuts the cost of the tokens those workloads use most.

The other headline is the KV cache. FP4 main-KV caching and a new attention scheme, Compressed Sparse Attention 2, bring the global cache down to 890 bytes per token, which is about a quarter of V4-Flash's. A technique called SWA Bounded Replay cuts the persistent cache DeepSeek stores on SSD to about one-eighth. That is why DeepSeek could cut cache-hit prices so far. The model also reads images natively (the vision encoder is trained from scratch), so the separate V4-Flash-Vision-Exp model is no longer needed.

The parameter counts can confuse people. DeepSeek's headline 552B is the backbone. The Hugging Face API counts about 763B parameters in the checkpoint because the 196B Engram lookup tables are stored alongside it. That difference matters a lot when you size hardware, which is covered below.

How much does DeepSeek V4.1 Flash cost?

The new prices took effect at 04:00 UTC on 10 September 2026. Like V4, DeepSeek bills by time of day: off-peak costs exactly half of peak. Rates below are from the official pricing page and the launch price card.

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour is off-peak, including whole weekends. That works out to about 79% of the week at the lower rate. Our DeepSeek peak and off-peak timezone guide converts those windows into eight local time zones.

Two numbers stand out. First, the cache-hit price is $0.006 per million at peak, 1/50th of the uncached rate. Agent loops that resend a large, stable prefix pay close to nothing for it. Artificial Analysis puts V4.1-Flash's blended price at $0.18 per million tokens (7:2:1 cache-hit/input/output mix), against $0.69 for V4-Pro 0813. Second, V4.1-Flash is roughly 3–7x cheaper than V4-Pro depending on token type, and it scores higher on most of DeepSeek's own agentic benchmarks. For the wider cost picture, see our DeepSeek V4 Flash vs Claude Opus 5 cost breakdown. V4.1-Flash is cheaper again than the V4-Flash prices in that analysis.

How do you call DeepSeek V4.1 Flash from the API?

Set the model to deepseek-flash. The base URL is unchanged: https://api.deepseek.com for the OpenAI format, or https://api.deepseek.com/anthropic if you want Claude-compatible tooling to talk to it. Thinking mode is on by default at high effort. You can set reasoning_effort to low, high or max. Other values are mapped: medium and xhigh become high, and ultra becomes max.

from openai import OpenAI

client = OpenAI(api_key="<DeepSeek API Key>", base_url="https://api.deepseek.com")

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Refactor this function to be iterative."}],
    reasoning_effort="max",
    extra_body={"thinking": {"type": "enabled"}},
)

print(response.choices[0].message.reasoning_content)  # chain of thought
print(response.choices[0].message.content)            # final answer

Some thinking-mode behaviours that catch people out. temperature, presence_penalty and frequency_penalty are silently ignored while thinking is on. top_p only takes values from 0.95 to 1.0. If your request includes tools, send the previous turns' reasoning_content back, or tool-calling quality drops. FIM completion only works in non-thinking mode. JSON output, tool calls, the Responses API and the Anthropic API format are all supported. For the full client setup, see our DeepSeek V4 API developer guide. Everything in it carries over once you change the model string.

Is DeepSeek V4-Pro deprecated?

No, not as of 22 September 2026. Here is what happened, in order:

1. 10 September. The launch announcement said "We're phasing out V4-Pro" and "Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches."
2. Before 14 September. DeepSeek reversed the decision. The official change log now says: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged. We will provide further notice should there be any changes."
3. Now. The pricing page lists deepseek-v4-pro as its own model, version DeepSeek-V4-Pro-0813, at its own (higher) rates, with no vision support.

So deepseek-v4-pro still reaches the real 1.6T-parameter V4-Pro 0813 and still bills at Pro prices. Those two facts matter together. If you stayed on Pro expecting the routing to cut your bill, it has not. You are paying roughly 4x for a model that DeepSeek's own benchmarks mostly rank below V4.1-Flash. For details on that checkpoint, see our DeepSeek V4-Pro 0813 benchmarks guide.

The legacy Flash names were changed. deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the models behind them are retired. Those requests now go to V4.1-Flash and bill at Flash prices. If you pinned deepseek-v4-flash, you have been running a different model since 10 September without changing any code.

Should you migrate from V4-Pro to V4.1-Flash?

For most workloads, yes, but test first. The "continue until further notice" wording means Pro could still be switched off at short notice. Hacker News commenters made the obvious point that silent model swaps behind an existing identifier break regression testing. A practical sequence:

• Record a Pro baseline now on your own eval set, while deepseek-v4-pro still serves the real model.
• Run the same set on deepseek-flash at reasoning_effort="max", then again at high. Effort changes both cost and latency.
• Check the areas where Pro still leads. On DeepSeek's own base-model table, V4-Pro is clearly ahead on factual recall (SimpleQA-Verified 55.2 vs 42.3), long context (LongBench-V2 51.5 vs 45.2) and multilingual knowledge (MultiLoKo 50.9 vs 45.5). For retrieval-light question answering over broad world knowledge, Pro may still be the better model.
• Pin the new model name explicitly. Use deepseek-flash, not the legacy alias, so the next retirement doesn't surprise you.

For coding agents, tool use and anything input-heavy, the cost difference and DeepSeek's agentic results both favour moving.

How good is DeepSeek V4.1 Flash on benchmarks?

All numbers in the table below are vendor-reported, from the Hugging Face model card, at maximum reasoning effort using DeepSeek's own "DeepSeek Harness" (DSH Minimal mode) for code-agent tasks. The comparison columns are DeepSeek's measurements of the other models, not those vendors' published figures.

† Text-only HLE subset. * The model card says 64.0; DeepSeek's change log says 65.4 for the same benchmark.

Is the Terminal-Bench 2.1 score of 90.6 real?

It is a real vendor-published number. It appears on the model card and in the official change log. It is not independently confirmed, and it is a poor guide to how the model will behave in your own terminal agent. Three reasons:

• Independent runs are much lower. Vals AI, using a neutral harness over three full trials, reports 74.53% on Terminal-Bench 2.1. That is 16 points below DeepSeek's figure, though still enough to make it the #2 open-weight model on Vals' run. We saw the same pattern with V4-Pro 0813: DeepSeek reported 87.9, while Vals measured 54.68% on the reference harness.
• The harness decides the score. DeepSeek's own scaffold table shows the same model scoring 90.6 on DSH Minimal, 88.0 in Claude Code, 86.1 in Pi, 85.0 in OpenCode and 84.1 in Codex. Even that range is closer together than neutral testing suggests.
• The benchmark is close to saturated. Every frontier model in DeepSeek's table scores 87.9–90.6 on Terminal-Bench 2.1. On the harder Terminal-Bench 3.0 and 4.0, the gap opens: V4.1-Flash scores 30.0 and 31.2, against Opus 5.0's 43.3 and 51.8. Those are more useful numbers for comparing models, and on them V4.1-Flash is behind the frontier, though well ahead of V4-Pro.

Read 90.6 as "V4.1-Flash is excellent inside DeepSeek's own harness", not as "V4.1-Flash is the best terminal agent available".

What do independent leaderboards say?

Third-party results broadly support the claim that V4.1-Flash beats V4-Pro, and do not support the claim that it matches the frontier.

• Artificial Analysis Intelligence Index v4.3.2: V4.1-Flash (max effort) scores 39, #7 among 114 open-weight models. On the same index version, V4-Pro 0813 scores 36, Kimi K3 scores 44, and Qwen3.8 Max (0902) scores 45. AA rebased the index in later versions, so do not compare these with older figures like V4-Pro's 53 on an earlier version.
• Speed: AA measures 218.9 output tokens/second and 0.99s time to first token, against 67.1 tok/s for V4-Pro 0813 and 38.2 tok/s for Kimi K3. The low active-parameter count shows up directly as speed.
• Verbosity: AA also flags it as "very verbose": 250M output tokens to complete its eval suite, against a 140M median. A low per-token price partly offsets a high token count, so measure cost per task, not per token.
• Vals Index: 57.86% (rank 17/60), which Vals calls the new #1 open-weight model on its index, at $0.303 per test.

Can you run DeepSeek V4.1 Flash locally?

Only with serious hardware, and not with the tools most people try first. "Flash" describes the active parameter count. Memory requirements depend on the total, and the total is large.

Hardware requirements

The Engram tables account for much of the footprint: roughly 183–189 GiB of FP8 lookup tables that are read sparsely. Runtimes that stream them from fast NVMe instead of keeping them in memory, such as DwarfStar's SSD streaming and vLLM's Engram CPU offload, are how anyone runs this outside a datacenter. The 1M-token KV cache, by contrast, is small at 890 bytes per token (under 1 GB for a full million tokens). For planning GPU purchases in general, our DeepSeek V4 VRAM and GPU requirements guide covers the V4 generation, and the self-hosting LLMs pillar guide covers serving stacks more broadly.

vLLM

V4.1-Flash is a new architecture (DeepseekV41ForCausalLM), so older vLLM builds that serve V4 will not load it. The official vLLM recipe requires vLLM 0.30.0 or later (NVIDIA nightly image from 10 September 2026 onward, or the deepseekv41-flash-0909 tag). A text-only serve command, adapted from the recipe's baseline:

vllm serve deepseek-ai/DeepSeek-V4.1-Flash \
  --tensor-parallel-size 4 \
  --language-model-only \
  --tokenizer-mode deepseek_v41 \
  --tool-call-parser deepseek_v41 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v41 \
  --max-model-len 1048576 \
  --max-num-seqs 128 \
  --max-num-batched-tokens 16384

The recipe's baseline also includes an AMD-specific --moe-backend aiter flag, which is left out above. On H200 it enables Engram CPU offload. DSpark speculative decoding is the only supported speculative method, configured through --speculative-config with 5 speculative tokens. Check the recipe page for your exact hardware before copying flags.

SGLang

NVIDIA's NVFP4 checkpoint card documents a tested SGLang setup on the lmsysorg/sglang:dev-cu13-dsv41 image across four GB300s:

python -m sglang.launch_server \
  --model-path nvidia/DeepSeek-V4.1-Flash-NVFP4 \
  --tp 4 \
  --context-length 1048576 \
  --reasoning-parser deepseek-v41 \
  --tool-call-parser deepseekv41 \
  --chunked-prefill-size 4096 \
  --max-running-requests 16

In that build, thinking is off by default. Pass a request-level reasoning_effort such as "max" to turn it on.

llama.cpp, Ollama and LM Studio

This is where most "run it locally" tutorials go wrong.

• llama.cpp: Upstream does not run V4.1-Flash yet. The conversion PR, ggml-org/llama.cpp#28696, opened on 10 September and is still a draft. The community GGUFs on Hugging Face (vcruz305/DeepSeek-V4.1-Flash-GGUF, for example) state in their own card: "These files do not run on upstream llama.cpp yet." They need a runtime fork where sparse attention is still being finished. Mainline llama.cpp support for V4-Flash does not cover V4.1.
• Ollama: The library entry is deepseek-v4.1-flash:cloud only. ollama run deepseek-v4.1-flash:cloud sends your prompts to Ollama's servers. It does not run on your machine. An open issue (#18379) asks for a downloadable build, with no maintainer response yet.
• LM Studio / MLX: Community MLX quants exist, but we could not confirm that any mainstream desktop app runs them. The only Apple Silicon route we could verify is antirez's DwarfStar (ds4, ds4.1flash branch), which ships calibrated V4.1 GGUFs:

# from a DwarfStar checkout with V4.1 support
./download_model.sh ds41f-q2
./ds4 -m gguf/DeepSeek-V4.1-Flash-Q2.gguf --ssd-streaming --ctx 32768

Those files only work with DwarfStar. They are not interchangeable with standard llama.cpp GGUFs or with V4-Flash weights. Expect single-digit tokens per second when streaming from SSD. If you want a local DeepSeek that works with standard tools today, V4-Flash is still the practical choice, as covered in our V4-Flash local setup guide. Note that its API alias now points to V4.1.

How does V4.1-Flash compare with V4-Pro, Qwen 3.8 and Kimi K3?

vs V4-Pro: V4.1-Flash wins on almost everything that matters for agents: DeepSWE (74.2 vs 62.7), Terminal-Bench 3.0/4.0 (30.0/31.2 vs 11.8/12.4), AutomationBench, speed and price. Pro keeps the lead on factual recall and long-context reading. Unless your workload is mainly knowledge-heavy Q&A, Flash is the better default.

vs Qwen 3.8: Qwen3.8 Max scores 6 points higher on AA's index but costs about 7x more on input and 5x more on output at DeepSeek's peak rates, and runs at under a fifth of the speed. Qwen3.8-Flash-Next is the real price competitor: cheaper on output ($0.47 vs $0.60–$1.20) with fewer active parameters. DeepSeek does not publish head-to-head numbers against it, and we found no neutral comparison yet, so test both on your own tasks. Our Qwen 3.8 lineup guide explains which Qwen variant to use.

vs Kimi K3: K3 is 5 points ahead on AA's index and 6.7 points ahead on HLE without tools in DeepSeek's own table. V4.1-Flash leads on DeepSWE (74.2 vs 67.5) and every Terminal-Bench version, runs almost 6x faster, and costs 10x less on input and more than 12x less on output. K3 is the smarter general reasoner. V4.1-Flash is far cheaper to run in an agent loop. See our Kimi K3 complete guide for more on K3.

Against closed frontier models such as Claude Opus 5.5, the harder agent benchmarks still show a clear gap. V4.1-Flash is not a frontier replacement. It is by far the cheapest model that gets reasonably close. More context on the DeepSeek line is in our DeepSeek V4 complete guide.

Who should use DeepSeek V4.1 Flash?

• Teams running high-volume agent loops: the low cache-hit price and fast prefill were built for workloads that resend large context on every step.
• Anyone currently on deepseek-v4-flash: you already use it through the alias, so change the model string to deepseek-flash and re-baseline.
• V4-Pro users: run the evals now, while you still control the timing of the switch.
• Self-hosters with hobbyist hardware: wait. Upstream llama.cpp and Ollama local support are not ready, and the realistic minimum is a 128 GB Mac with fast SSD streaming or a multi-GPU datacenter node.

Teams that build their own agent infrastructure on top of open models like this one sometimes need extra engineering capacity to do the evaluation, serving and migration work. Codersera can help you hire vetted remote developers who have shipped LLM-backed systems.

FAQ

What is the API model name for DeepSeek V4.1 Flash?

deepseek-flash, at base URL https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format). The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp also go to V4.1-Flash now, but DeepSeek describes that routing as temporary.

Did DeepSeek deprecate V4-Pro on 14 September 2026?

No. DeepSeek announced on 10 September that deepseek-v4-pro would route to V4.1-Flash from 14 September, then reversed that "in response to user demand". V4-Pro 0813 is still served at its original prices until further notice. DeepSeek still describes V4.1-Flash as the stronger model.

How much does DeepSeek V4.1 Flash cost?

Per million tokens at peak: $0.006 cached input, $0.30 uncached input, $1.20 output. Off-peak is half: $0.003, $0.15 and $0.60. Peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays. Everything else, including weekends, is off-peak.

Is DeepSeek V4.1 Flash open source?

Yes. The code and weights are MIT-licensed on Hugging Face at deepseek-ai/DeepSeek-V4.1-Flash, with a technical report in the same repository. Commercial use is allowed.

Can I run DeepSeek V4.1 Flash with Ollama or llama.cpp?

Not locally, as of 22 September 2026. Ollama offers only a cloud tag, and upstream llama.cpp cannot yet run V4.1 GGUFs because the support PR is still a draft. vLLM 0.30+ and SGLang dev builds work on multi-GPU servers, and antirez's DwarfStar runs a Q2 build on a 128 GB Mac with SSD streaming.

Is the Terminal-Bench 2.1 score of 90.6 trustworthy?

It is DeepSeek's own number, measured in DeepSeek's own harness. Vals AI's independent run measured 74.53%. The benchmark is also close to saturated among frontier models, so Terminal-Bench 3.0 (30.0) and 4.0 (31.2) are better guides to real agent capability.

Sources

• DeepSeek: DeepSeek-V4.1-Flash release announcement (10 Sept 2026)
• DeepSeek API change log (V4-Pro continuation notice)
• DeepSeek API: Models & Pricing
• DeepSeek API: Thinking Mode
• Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash model card
• vLLM Recipes: DeepSeek-V4.1-Flash
• NVIDIA: DeepSeek-V4.1-Flash-NVFP4
• antirez: DeepSeek V4.1 Flash GGUF for DwarfStar
• vcruz305: DeepSeek-V4.1-Flash GGUF (status notes)
• Artificial Analysis: DeepSeek V4.1 Flash
• Vals AI: DeepSeek V4.1 Flash