DeepSeek V4.1-Flash vs MiniMax M3.1 Flash

Both are low-cost “Flash” models released in the same week: DeepSeek V4.1-Flash-1017 on October 17 and MiniMax M3.1 Flash on October 19. “Vendor” marks a claim from the model’s maker; “AA” marks an independent measurement by Artificial Analysis.

Side by side

	DeepSeek V4.1-Flash-1017	MiniMax M3.1 Flash
Input, $ per 1M tokens (cache miss)	$0.15 off-peak / $0.30 peak (DeepSeek)	$0.15 flat (MiniMax)
Output, $ per 1M tokens	$0.60 off-peak / $1.20 peak (DeepSeek)	$0.60 flat (MiniMax)
Cached input, $ per 1M tokens	$0.003 off-peak / $0.006 peak (DeepSeek)	$0.015 (MiniMax)
AA Intelligence Index	45 (AA)	43 on the live page (AA); 42 at launch (VentureBeat)
AA cost per Index task	$0.26 (AA)	$0.22 (AA)
Measured speed (AA)	208.7 tokens/s, TTFT 1.03 s	93.1 tokens/s, TTFT 1.39 s
Vendor benchmarks	DeepSWE v1.1: 77.0 · Automation-Bench v1.0.6: 60.5 · Agents’ Last Exam: 35.4 (HF card)	DeepSWE v1.1: 73.7% · Automation-Bench v1.0.6: 62.2 (MiniMax chart, via VentureBeat) · Agents’ Last Exam: 29.6 (MiniMax)
Independent tests (AA, via VentureBeat)	HLE 43% · CritPt 24% · AA-Omniscience accuracy 43%	HLE 48% · CritPt 22% · AA-Omniscience accuracy 56%, hallucination rate 93%
Context window	1M tokens (DeepSeek)	1,000,000 tokens (MiniMax)
Max output	384K tokens (DeepSeek)	Not in official docs
Weights / licence	Open weights, MIT (HF card)	Not released; announced for later (MiniMax)
Notable features	Thinking and non-thinking modes, JSON output, tool calls, vision, built-in Search & Fetch APIs	Tunable reasoning effort, interleaved thinking with tool use, text/image/video input, paid web search tool
Known limits	Peak hours double the price; very verbose (AA)	Thinking cannot be turned off; high hallucination rate (AA); slower than average (AA)

DeepSeek V4.1-Flash-1017

DeepSeek calls this release a “minor update” to V4.1-Flash. The API model name is deepseek-flash, and the legacy deepseek-v4-flash* names now route to it (DeepSeek pricing, release news).

• Peak pricing: 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Off-peak is half price (DeepSeek).
• OpenRouter: passes through the same time-of-day pricing (OpenRouter).
• API features: thinking mode (default) and non-thinking mode, JSON output, tool calls, Responses API, an Anthropic-format API, chat prefix completion (beta), FIM (beta, non-thinking only) and vision.
• Concurrency: limit of 2,500 (DeepSeek).
• Caching: AA lists a 98% cache discount (AA).
• Search & Fetch APIs: run on DeepSeek’s in-house search engine, $0.75 per 1,000 successful requests each. Server-side calls through DeepSeek Harness, Claude Code and Codex are free until billing starts on 2026-10-31 16:00 UTC (release news).
• Parameters: the sources conflict. Hugging Face shows 763B (HF card); AA lists 552B total and 16B active (AA).
• Weights: on Hugging Face and ModelScope under the MIT licence (release news).

Vendor benchmarks, from the model card:

• HLE without tools: 42.6 (44.0 on the text-only subset)
• Terminal-Bench 4.0: 42.3
• ProgramBench: 29.1
• NL2Repo: 68.2
• Apex Bench: 72.1
• DeepSWE cost: DeepSeek claims about $0.20 per task at 77.0 (The Decoder)

Independent results from AA:

• Intelligence Index: 45 at “max” reasoning effort
• Output tokens used on the Index: 210M (AA calls it “very verbose”)
• Time to first token: 1.03 s

MiniMax M3.1 Flash

MiniMax released M3.1 Flash on October 19. An early version ran anonymously as “Space Bunny” and processed more than 200T tokens (MiniMax).

Vendor claims, from the MiniMax blog:

• About 3× faster inference and 70% lower agentic usage cost than its predecessor
• Speed: 200+ tokens/s, P50 170+
• BankerToolBench: 76.2
• OSWorld V2: 61.7
• CyberGym: 82%
• #1 among Chinese models on OSWorld V2 and ALE (MiniMax’s claim)

Independent results from AA:

• Intelligence Index: 43 on the live page (42 at launch, per VentureBeat)
• Output tokens used on the Index: 180M on the live page (about 189M in the launch figures, per VentureBeat)
• Speed: 93.1 tokens/s, which AA calls “slower than average”
• Time to first token: 1.39 s
• AA-Omniscience: 56% accuracy, 93% hallucination rate (wrong answers as a share of non-correct responses), Omniscience Index 16 (per VentureBeat)

Features and access:

• Reasoning: tunable effort from low to max (default max). Thinking cannot be turned off; requests that try return a 400 error (MiniMax docs).
• Tools: interleaved thinking with tool use (MiniMax docs).
• APIs and inputs: Anthropic- and OpenAI-compatible endpoints; text, image and video input.
• Paid extras: web_search server tool at $0.01 per request; prompt-cache reads at $0.015 per 1M tokens (MiniMax pricing).
• Where to get it: MiniMax API, M Plan and MiniMax Code; a Pro model is promised later (MiniMax). AA lists one API provider.
• Weights: not released yet. MiniMax says they “will be released later” (MiniMax); licence terms have not been given (VentureBeat). AA marks the model proprietary with no parameter count disclosed.

Who each suits

DeepSeek V4.1-Flash suits teams that want:

• Self-hostable weights under the MIT licence
• The fastest measured API speed of the two
• Built-in web search through the Search & Fetch APIs
• The cheapest cache hits
• Low-cost batch work scheduled off-peak

MiniMax M3.1 Flash suits:

• Cost-sensitive agent and automation workloads, such as SaaS workflows and finance tool chains
• Jobs that need the same price at any hour
• Teams willing to add retrieval or verification for anything fact-sensitive, given the 93% hallucination rate