Ant Group Open-Sources Ling-3.1-flash Under MIT — Priced Under DeepSeek's Flash Tier

Verdict: Ant Group's Ling team open-sourced Ling-3.1-flash yesterday (October 15) under MIT — a 560B-parameter MoE with 25B active per token and a 1M-token context window. The API prices at $0.30 per 1M input and $0.90 per 1M output, which undercuts DeepSeek's flash tier on output tokens ($1.20 peak) with no peak/off-peak scheduling required. But I'd hold off on switching anything on price alone: Artificial Analysis flagged this model as very verbose, and verbosity is exactly the kind of thing that eats a 25% paper discount.

What actually shipped

The weights are on Hugging Face under the inclusionAI org, MIT licensed — commercial use included. The architecture is a hybrid-linear MoE: ~560B total parameters, ~25B active per token. On Artificial Analysis, it scores 41 on the Intelligence Index (#6 out of 119) and runs at 215 output tokens per second (#3 out of 77) — fast, and smart enough to sit in the same conversation as frontier-adjacent closed models.

The official framing from Ant is "agent" work: office tasks, long-horizon software development, complex multi-step jobs. The timing isn't accidental — Meta shipped its Muse Wonder agent models on October 12 and Google pushed Gemini Agent and its Gemini 4 Argon flagship days before. Every lab is now selling the same story: AI that finishes real work, not just chats. (Chinese coverage of this push — and Ant's response — came via 智东西/Zhidongxi, October 15.)

The pricing math against DeepSeek

Here's the part most coverage will skip. DeepSeek's current flash tier — DeepSeek-V4.1-Flash, served when you call deepseek-flash — also has a 1M context window, so the headline context number is not a differentiator here. What is different is how you pay:

	Ling-3.1-flash (official API)	deepseek-flash (DeepSeek V4.1-Flash)
Input, cache miss	$0.30 flat	$0.30 peak / $0.15 off-peak
Output	$0.90 flat	$1.20 peak / $0.60 off-peak
Context	1M	1M
License	MIT weights	Closed, API only

On flat pricing, Ling wins the output column by 25% and ties on input. If you can route batch work into DeepSeek's off-peak windows (01:00–04:00 and 06:00–10:00 UTC, weekdays), DeepSeek's $0.60 output rate flips the table. So the honest answer to "is Ling cheaper than DeepSeek?" is: only if you can't or won't schedule around peak hours. We keep a running comparison of the full provider field in our cheapest DeepSeek API breakdown.

The catch: verbosity eats discounts

Artificial Analysis measured Ling-3.1-flash generating about 220M tokens per Intelligence Index task versus a 140M median — roughly 50% more output than the typical model for the same work. Their own summary calls it "very verbose." Their blended rate (7:2:1 cache-hit/input/output) comes out to $0.19 per 1M tokens, which is the number that will get screenshotted. My take: that number is marketing-adjacent. If the model emits ~50% more tokens to do the same job, your effective cost per task rises by the same factor — and a verbose model at $0.90/M output can cost more in practice than a terse one at $1.20/M.

I haven't run the model myself, so this is arithmetic against AA's published numbers, not a benchmark of my own. But the direction is what I'd bet on: test Ling against your own workload before migrating anything, and measure tokens-per-task, not price-per-token. This is the same trap behind "off-peak pricing looks 50% cheaper" — the sticker rate is never the whole cost. (We covered the DeepSeek side of that in DeepSeek off-peak pricing explained.)

Self-hosting: open weights, heavy lift

MIT weights mean anyone can fine-tune and serve Ling-3.1-flash commercially — that's real, and it's the one advantage DeepSeek's closed flash tier can't match at any price. But keep expectations grounded: 560B total parameters is a datacenter deployment, not a weekend build. If your goal is running a capable model on your own hardware, DeepSeek's smaller flash models remain the practical entry point — see our DeepSeek 4.1 Flash deployment guide. Ling's 25B active parameter count helps serving throughput, but the full weight set still has to live in memory.

For the head-to-head against the other Chinese price fighter, our Qwen3.8 Flash vs DeepSeek 4.1 Flash comparison covers the pricing details line by line — Ling now sits in that same weight class.

What I'd watch next

Two things will decide whether Ling-3.1-flash actually gets adopted outside China: whether third-party hosts (OpenRouter and friends) keep the price near the official $0.30/$0.90, and whether the verbosity shows up in real agent workloads or just in benchmark answer-length habits. If a future checkpoint trims token output while holding the Intelligence Index score, this becomes the default "cheap and open" pick. Until then, it's a strong number-two to watch, not a switch-today call.

Caveat (Oct 16, 2026): prices and scores are as published on October 15–16, 2026 — AA's figures are based on InclusionAI's official API pricing, and provider pricing may vary. DeepSeek peak/off-peak windows and rates per DeepSeek's official pricing page on the same date.