Ant’s Ling-3.0-flash-VL adds vision to a sparse 124B reasoning model

Ant Group’s InclusionAI lab has released Ling-3.0-flash-VL, an MIT-licensed, open-weight reasoning model that accepts text, images, and short video clips and produces text. The model activates 5.5 billion of its 124 billion parameters for each token and scored 25 on the Artificial Analysis Intelligence Index v4.3, placing it on the evaluator’s intelligence-versus-active-parameters Pareto frontier. Full results appear on the Artificial Analysis page.

Ling-3.0-flash-VL extends the text-focused Ling-3.0-flash architecture with visual inputs, a 256K-token context window, and self-hosting support. Its sparse activation can reduce per-token computation, but deployment still requires enough memory to load the complete checkpoint plus serving overhead.

124B weights, 5.5B active parameters

Ling-3.0-flash-VL specifications

Feature	Details
Architecture	Sparse mixture of experts
Parameters	124B total, 5.5B active per token
Expert routing	8 of 512 routed experts plus one shared expert
Context window	Up to 256K tokens
Inputs	Text, images, and short video
Output	Text, including generated reasoning traces
Weight formats	BF16 and FP8 available; FP4 and INT4 planned
Runtimes	SGLang and InclusionAI’s vLLM fork
License	MIT, permitting commercial use and modification with the required notice

A router sends each token through eight of the model’s 512 specialized experts and one shared expert. This mixture-of-experts design draws from a large parameter pool while limiting the portion used for each generation step.

The attention stack contains 35 Kimi Delta Attention layers and seven Gated Multi-head Latent Attention layers arranged in a 5:1 pattern. Kimi Delta Attention is a linear-attention design intended to reduce the memory growth associated with long sequences. The seven latent-attention layers retain more conventional attention behavior where the architecture needs it.

Sparse activation lowers arithmetic per token, while capacity planning remains tied to the complete checkpoint. The BF16 weights occupy about 255GB and the FP8 weights about 128GB before runtime memory, cache, batching, and framework overhead. Long-context requests increase those additional requirements, making multiple high-memory GPUs the practical baseline.

Why the benchmark number changed

Artificial Analysis assigns Ling-3.0-flash-VL a score of 25 on Intelligence Index v4.3. The evaluator reports a median of 8 for similarly sized open-weight models and places Ling on its active-parameter Pareto frontier. Within the compared set, obtaining a higher score requires activating more parameters per token.

Intelligence Index v4.3 comparison

Model or group	Score
Ling-3.0-flash-VL	25
Qwen3.5 122B A10B, reasoning mode	16
Mistral Medium 3.5	15
Similar-size open-weight median	8

An earlier vendor report gave the model a score of 42 on Intelligence Index v4.1.1, four points above the text-only Ling-3.0-flash. Protocol changes prevent a direct comparison between that result and the independent score of 25 on the harder v4.3 index.

Index v4.3 aggregates ten evaluations, including SciCode, Humanity’s Last Exam, Terminal-Bench, AA-Omniscience, and agentic workflow tests. The composite score covers broad reasoning and tool-use performance. Visual accuracy requires separate evaluation because the cited index result does not isolate image and video capabilities.

Fewer fabrications, faster generation

AA-Omniscience recorded 14% accuracy and a 22% hallucination rate for Ling-3.0-flash-VL. Inkling Small reached 33% accuracy but hallucinated on 63% of the evaluation. These figures indicate a more conservative response pattern from Ling on this benchmark, though teams should measure abstention and fabrication rates against their own domain data.

InclusionAI’s first-party endpoint produced 142.9 output tokens per second in Artificial Analysis testing, compared with a 99-token-per-second median for similarly sized open-weight models. Time to first token was 2.24 seconds. Hosted throughput will vary with prompt length, batching, provider load, hardware, and sampling settings.

Shell agents and token budgets lag

Agentic evaluations expose the model’s weakest area. Ling-3.0-flash-VL scored 16% on AutomationBench-AA, which tests guarded business-application workflows, and 0% on Terminal-Bench v4.0, a demanding terminal-use benchmark. Products centered on autonomous shell work or coding agents need stronger task-specific evidence before adoption.

Reasoning length also raises serving costs. The model generated roughly 50,000 output tokens per Intelligence Index task, compared with about 30,000 for Inkling Small at a similar overall score. Output-token caps, stopping rules, and reasoning budgets will therefore affect both latency and cost.

The text-only Ling-3.0-flash used about 240 million output tokens to complete its Intelligence Index run at a reported total cost of $73. That historical figure describes a different model and evaluation run, so it cannot establish the VL release’s production cost.

Server-class deployment is required

The release offers two self-hosted checkpoints through the InclusionAI repository: BF16 at about 255GB and FP8 at about 128GB. InclusionAI lists SGLang and its published vLLM fork as supported serving paths. Planned FP4 and INT4 quantizations could lower memory requirements, although no release date is specified.

InclusionAI recommends a temperature of 0.6, top_p of 0.95, and top_k of 20. Those values provide a starting point; production deployments should retune them for answer length, determinism, hallucination rate, and tool-call reliability.

Hosted access is available through Ant’s Ling platform API. At publication, Novita AI also offered time-limited free hosting with provider limits of 262K input tokens and 32K output tokens. Endpoint limits can differ from the model’s advertised context window and should be verified before integration.

A production price for the VL model was not included in the reported material. The text-only predecessor cost $0.075 per million input tokens and $0.22 per million output tokens on the first-party API, which provides historical context rather than a reliable VL estimate.

Ant brings vision into the Ling family

Ant’s earlier model portfolio assigned multimodal work to the Ming line. Ling-3.0-flash focused on text, tools, coding, and research. Ling-3.0-flash-VL brings visual input into the sparse Ling reasoning architecture, giving developers one model for long text, document images, screenshots, figures, and short clips.

The broader Ling family now includes Ling-3.0-tiny for local deployment, Ling-3.0-flash-VL for visual inputs, and Ling-3.0-flash-Sante for healthcare and life-science workloads. This variant strategy adapts the family to different infrastructure and evaluation requirements. The VL release’s MIT license also allows teams to modify and redistribute its weights under the license terms.

Where the model fits

Ling-3.0-flash-VL best matches teams that have server-class GPU capacity and one or more of these requirements:

• Multimodal reasoning across long documents containing text, images, or short video clips
• Self-hosting for data residency, privacy, customization, or predictable infrastructure control
• Retrieval-augmented generation pipelines that need conservative responses and explicit abstention
• Large expert capacity with relatively low active computation per generated token

Terminal-heavy coding agents are a poor fit based on the reported 0% Terminal-Bench result. Cost-sensitive reasoning pipelines also require strict output controls because the model generates substantially more tokens than some similarly scoring peers.

Checks before production

• Evaluate image and video accuracy on representative inputs, since the composite index does not isolate visual performance.
• Measure peak memory and latency at the intended context length, concurrency, and precision.
• Set reasoning-token budgets, stopping rules, and request-level cost limits.
• Reproduce tool-use and agent benchmarks with the exact runtime, prompts, and integrations planned for deployment.
• Confirm hosted pricing, rate limits, retention policies, and endpoint context limits with the provider.