
by InclusionAI
The successor to Ling 3.0 Flash in InclusionAI's flash line: a much larger sparse model tuned for agents, search and office tasks, offered through hosted APIs while the weights stay unreleased.
Ling 3.1 Flash, InclusionAI's September 2026 release, scored 41 on the Artificial Analysis Intelligence Index, roughly double its predecessor's 20. It suits teams prototyping coding and tool-using agents on a free trial, but anyone needing downloadable weights, vision input or published privacy terms should wait or choose a rival.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0090 | $0.0009 | $0.0099 |
| Support reply | $0.0006 | $0.0003 | $0.0009 |
| One coding agent run | $0.060 | $0.018 | $0.078 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Artificial Analysis recorded a rate of $0.30 per million input tokens and $0.90 per million output tokens, with cached input at $0.06 (an 80% discount). The launch trial is free: Novita AI, OpenRouter and Vercel AI Gateway all showed $0 on 8 October 2026, and InclusionAI described the free period as two weeks, after which the service turns paid. At the listed rate a request of one million input and one million output tokens costs $1.20. GLM-5.3 Flash and DeepSeek V4.1 Flash list lower rates on their Hokai records, so the case for Ling is speed and agent behavior, not price. The model is verbose, which pushes the real bill above a per-token comparison.
Ling 3.1 Flash is a 560-billion-parameter mixture-of-experts reasoning model from InclusionAI, Ant Group's AI lab, that activates about 25 billion parameters per token. It accepts and returns text only, and its trial hosting serves a 262,144-token context while the weights remain unreleased.
Reads
Writes
Sparse mixture of experts
Only about 4.5% of the weights run for each token, the design choice behind its measured speed.
Fast output
Artificial Analysis measured a median 211.4 output tokens per second on InclusionAI's own API.
Reasoning on by default, switchable
Thinking runs unless you turn it off with the reasoning parameter; Artificial Analysis saw about 9.5 seconds of reasoning before the first answer token on its long-prompt test.
Terminal and tool work
Function calling with tool_choice works on Novita AI and OpenRouter, and the Terminal-Bench 4.0 run scored 33.3%.
Long-context reasoning
It scored 83% on AA-LCR, Artificial Analysis's long-context reasoning test.
Anthropic-style endpoint
Novita lists an Anthropic messages endpoint beside chat completions, so Claude-format clients can point at it.
Artificial Analysis Intelligence Index v4.3.2
Source: Artificial Analysis model page, opened 2026-10-08Cheaper than 75% of the 107 GA models with a published price, and rank 14 of 60 on output speed as cited from Artificial Analysis. Ranked against GA models; this record is not GA.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.45/M | #28 / 107 | peer median $1.71/M |
| Output speedtokens/s | 211 tok/s | #14 / 60 | peer median 95 tok/s |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| InclusionAI | 262K | $0.30 / $0.90 | text → text | |
| Z.ai | 1M | $0.15 / $0.50 | text, image, video, pdf → text | |
| DeepSeek | 1M | $0.15 / $0.60 | text, image → text | |
| 1M | $0.75 / $3.75 | text, image, video, audio, pdf → text |
| Name | Endpoint | Auth |
|---|---|---|
| Novita AI | https://novita.ai/models/model-detail/inclusionai-ling-3.1-flash | API Key |
| OpenRouter | https://openrouter.ai/inclusionai/ling-3.1-flash | API Key |
| Vercel AI Gateway | https://vercel.com/ai-gateway/models/ling-3.1-flash-free | API Key |
Reasoning runs before every answer
Thinking is on by default. Artificial Analysis recorded about 9.5 seconds of reasoning before the first answer token on its long-prompt test, on top of the first chunk arriving after 1.79 seconds.
Workaround: Switch the reasoning parameter off on hosts that allow it for latency-sensitive calls, and stream responses so users see progress.
Very verbose output
Artificial Analysis counted more than twice the median output tokens on its index run, which inflates cost per task on any per-token plan.
Workaround: Set max_tokens well below the 32,768 ceiling, ask for terse formats, and track output tokens per task from the first day.
Context ceiling differs by source
Artificial Analysis lists a 1M-token window, but the hosted trial endpoints serve 262,144 tokens until InclusionAI enables the larger window after the trial.
Workaround: Plan and test against 262,144 tokens, and treat the 1M figure as a future capability until a host confirms it.
Free price is temporary
Hosts show $0 per million tokens only for the launch trial, and InclusionAI has said the service turns paid afterwards.
Workaround: Budget at the rate Artificial Analysis recorded (see the price notes) before wiring it into a production pipeline, and re-check the host's price page weekly.
Weights promised, not shipped
InclusionAI says it will open-source the model after the trial but has named no date or license, and the expected Hugging Face repository did not exist on 8 October 2026.
Workaround: Do not design around self-hosting yet; re-check the InclusionAI Hugging Face organization page for a Ling-3.1-flash repository before committing.
Ling-3.1-flash · 30 Sep 2026
Successor to Ling 3.0 Flash: 560B total and ~25B active parameters, hybrid reasoning, text only, two-week free trial at 256K context, open-source release and 1M context promised after the trial.
Ling-3.0-flash · 24 Jul 2026
Predecessor: 124B total and ~5.1B active parameters, 256K context; its weights are public on Hugging Face.
Artificial Analysis records $0.30 per million input tokens, $0.90 per million output tokens and $0.06 for cached input, measured on InclusionAI's own API. During the launch trial Novita AI, OpenRouter and Vercel AI Gateway charge nothing, so a test run cost $0 on 8 October 2026. Against Claude Haiku 5.5's Hokai-listed $0.10 and $0.50, the list rate is three times higher on input and almost twice as high on output.
On the numbers Hokai holds, no: GLM-5.3 Flash scores slightly higher on the Artificial Analysis index, lists a lower per-token rate, accepts images, video and PDFs, and ships MIT-licensed weights you can download. Ling's case rests on its measured output speed, its agent-oriented tuning and the free trial. Pick GLM unless you have a specific reason to test Ling.
Not yet. InclusionAI says it plans to release the weights after the free trial, but no license, date or checkpoint format has been published, and the inclusionAI/Ling-3.1-flash page on Hugging Face returned a 404 on 8 October 2026. Artificial Analysis currently labels it proprietary, so treat it as an API-only model until a repository appears.
InclusionAI has not published a retention or training policy for this model that we could find. When you call it through Novita AI, OpenRouter or Vercel AI Gateway, that host's own terms apply to your prompts, so read them before sending customer or proprietary data. Until a policy exists, keep regulated or confidential material out of the trial.
Developers building terminal or tool-calling agents who can cap output length and want a no-cost test of a large reasoning model. Pass on it if you need vision input, a self-hosted checkpoint or a signed data agreement, and look at GLM-5.3 Flash or Gemini 3.8 Flash instead. Its verbosity also makes it a poor fit for long-form chat where token spend is the main cost.
InclusionAI, the AI lab inside Ant Group (see InclusionAI), announced Ling-3.1-flash on 30 September 2026, China time. It follows Ling 3.0 Flash from July, which had 124 billion total parameters and about 5.1 billion active, and it sits on a different branch from the finance fine-tune Ling 3.0 Flash Fin. The new model is a sparse mixture of experts with roughly 560 billion total and 25 billion active parameters. InclusionAI positions it for agent tasks, search, office software and specialist applications. The lab's other releases are collected on the InclusionAI provider page.
The independent evidence comes from Artificial Analysis, which ran its ten-evaluation Intelligence Index and scored the model 41. Inside that index it reached 33.3 percent on Terminal-Bench 4.0, 39.4 percent on Humanity's Last Exam and 54.1 percent on SciCode, plus an Elo of 1,622 on GDPval-AA, its test of real knowledge-work deliverables. On AA-Omniscience it answered 29.1 percent of questions correctly and recorded a 37.9 percent hallucination rate, Artificial Analysis's measure of how often a model gives a wrong answer instead of abstaining when it does not know. Third-party write-ups quote higher vendor figures taken from launch graphics we could not inspect, so every score on this page comes from the Artificial Analysis run.
On InclusionAI's own API, Artificial Analysis measured a median 1.79 seconds to the first chunk and about 13.6 seconds end to end on its long-prompt test, of which roughly 9.5 seconds was reasoning before the answer began. The index run consumed about 218 million output tokens, more than twice the 100 million median Artificial Analysis cites for comparable models, and cost about 99 cents per task at the listed rate.
The hosted trial endpoints serve a 256K-token window and up to 32,768 output tokens. InclusionAI says the full 1M-token window arrives when the trial ends and the service turns paid. Input and output are text only, which places it among the text-only models in the directory. Novita AI and OpenRouter expose function calling with tool_choice and a reasoning parameter that is on by default, and Novita adds an Anthropic-style messages endpoint beside chat completions.
On 8 October 2026 the model was free to call on Novita AI, OpenRouter and Vercel AI Gateway. Those are promotional rates, not the list price, so the trial is the place to test before billing starts. The same hosts appear in the AI gateways category, and the AI agent builders category covers the frameworks that would sit in front of an agent model like this one. Compare it with GLM-5.3 Flash, DeepSeek V4.1 Flash, Gemini 3.8 Flash and Claude Haiku 5.5 before committing. On Hokai's own records GLM-5.3 Flash lists a lower rate, a slightly higher index score and MIT-licensed weights; DeepSeek V4.1 Flash lists a lower rate with MIT weights and a slightly lower score; Gemini 3.8 Flash scores about the same, costs more and takes image, audio and video input; Claude Haiku 5.5 lists a lower rate and a higher score. Further up the index, DeepSeek V4 Flash lists a lower rate and a higher score with open weights, Tencent's Hy3 scores far higher at a similarly low rate, and Kimi K3 scores higher at a much higher rate. The model leaderboard sorts all of them on the same index.
InclusionAI says it plans to open-source the model after the trial but has not named a license, a date or a checkpoint format. On 8 October 2026 the expected Hugging Face page at inclusionAI/Ling-3.1-flash returned a 404, Artificial Analysis classes the model as proprietary, Hokai files it as a preview release while the trial runs, its hosted context sits in the 128K to 400K bucket, and we found no system card, data-retention statement or training-data summary. Ling 3.0 Flash and its Fin fine-tune both have public Hugging Face repositories, and the Fin model carries the MIT license, which is a pattern, not a promise.
The sensible use today is a time-boxed trial: run your own terminal and tool-calling tasks while the hosts are free, cap output length because the model is verbose, and decide on the list rate afterwards. Teams that need to self-host, send images or sign a data-processing agreement should look at the peers above until InclusionAI publishes weights and terms.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from InclusionAI's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.