Ling-3.0-flash-VL vs Ling 3.0 Flash: One 124B Brain, Now With Eyes

Ling-3.0-flash-VL and Ling 3.0 Flash are not rivals, and reading this pairing as a model shootout will send you to the wrong conclusion. Ling-3.0-flash-VL is the vision-language checkpoint that Ant Group's inclusionAI lab released this week — weights live on Hugging Face under an MIT license since September 4, 2026 — and it is built directly on Ling 3.0 Flash, the lab's 124B-parameter open-weight text model that has anchored its fast tier since late July. The two share the same hybrid-linear MoE design, the same sparse-activation economics, the same 256K-context serving recipe, and the same MIT license. What the VL checkpoint adds is the one thing the text model never had: native image and video input, carried through a ViT encoder, a two-layer MLP projector, and VideoRoPE temporal encoding. So the comparison reduces to a single practical question — if your workload never looks at a picture, does the model with eyes earn the switch?

The reason the question is worth asking at all is that inclusionAI does not frame Ling-3.0-flash-VL as a separate product line. Its model card calls it "our next-generation native multimodal model," built on Ling-3.0-flash, inheriting that model's language, reasoning, and long-context ability and extending them with vision that participates in the full loop of understanding, reasoning, acting, and verification — not vision as a bolt-on input. For a model family whose earlier flagship release was explicitly text-only, that is a real shift, and it changes which checkpoint a text-and-tools team should standardize on.

Two checkpoints, one backbone

Ling 3.0 Flash, the opponent in this comparison, is the mature one of the two. Announced in late July 2026 and open-sourced under MIT on August 7, it is a hybrid-linear mixture-of-experts model with 124B total parameters and roughly 5.1B active per token — 35 KDA layers and 7 Gated MLA layers stacked at a 5:1 ratio, 512 routed experts with 8 activated, a 256K native context served at 262,144 tokens. It is text-only, with tool-calling support, thinking enabled by default through the chat template, and an unusually low cost profile for its size. It is also the documented half of the pair: Artificial Analysis lists it on its live leaderboard, where it is ranked first among the 63 models in its class, and has measured output at roughly 313 tokens per second on the vendor API.

Ling-3.0-flash-VL keeps that architecture and adds a visual stack on top. The card describes a ViT visual encoder whose features are aligned into the text space by a two-layer MLP projector, with VideoRoPE encoding both spatial position and temporal order so the model can do event localization and long-video reasoning. The backbone is the same 5:1 KDA-to-MLA hybrid, this time at 124B total / 5.5B active per token, with a 42-layer trunk. Inputs are text, image, and video; output is text. Context is advertised up to 1M tokens, with the recommended SGLang recipe serving 256K via YaRN scaling. Like its text sibling it ships thinking-on by default (disable per request with chat_template_kwargs), and it runs under either SGLang or a custom inclusionAI vLLM fork. The BF16 release is about 250 GB on disk across 64 safetensor shards — the same quarter-terabyte class as the text model's weights.

How fresh is it? At the time of writing the VL repository has a download count in the dozens, its model card was still being updated, and Artificial Analysis has not listed it. Everything below that is not explicitly attributed to Artificial Analysis is inclusionAI's own reporting.

The scoreboard

The asymmetry here is not quality — it is modality. Six dimensions capture the decision:

• Intelligence (vendor card, AA Index v4.1.1) — Ling-3.0-flash-VL: 42, vendor-reported. Ling 3.0 Flash: 38, the earlier index figure the card cites.

• AA live board today (v4.3) — Ling-3.0-flash-VL: not listed yet. Ling 3.0 Flash: estimated 25, ranked #1 of 63 in its class.

• Inputs — Ling-3.0-flash-VL: text, image, and video. Ling 3.0 Flash: text only.

• Context — both claim up to 1M tokens; both are practically served at 256K (262,144) today.

• Price — Ling-3.0-flash-VL: no list price yet; MIT open weights only. Ling 3.0 Flash: about $0.07 per 1M input and $0.22 per 1M output on the vendor's first-party API.

• Pick it for — Ling-3.0-flash-VL: documents, screenshots, charts, video, and GUI-acting agents. Ling 3.0 Flash: text-heavy tool-calling and long-horizon agentic pipelines.

The scoreboard the text model can't enter

This is where the two actually diverge, and the divergence is structural rather than competitive: on image and video benchmarks the text-only Ling 3.0 Flash has no entry at all, because it cannot accept the input. Ling-3.0-flash-VL's own chart — inclusionAI's evaluation, unreproduced, run partly in-house and partly against the competitors' APIs under official test settings — posts numbers across the three buckets the card emphasizes. On complex-image understanding it shows an MMMU-Pro of 79.0 and a MathVision of 84.87, with a much lower 19.88 on Humanity's Last Exam-Multimodal reflecting how hard that set is for everyone. On document and chart work it is strong: OmniDocBench1.5 at 91.35 and CharXiv_RQ at 81.30. And on the agentic-multimodal suites that make this release interesting — ClawEval-MM at 59.9, WebVoyager at 90.83, Vision2Web at 57.69 — it is asked to read an interface and act on it, which is precisely the GUI-agent job the text model cannot take.

Read those in context and a coherent picture emerges: this is not a model tuned to win trivia over images, it is a model tuned to turn pixels into action — read the layout, follow the chart, operate the page. On the vendor's own chart it leads the three-way comparison field (Qwen3.8-27B at high reasoning effort, Gemini 3.5 Flash-Lite, and Kimi-K2.6) on OmniDocBench, WebVoyager, and ClawEval-MM, and trails on hard knowledge-and-reasoning sets like MathVision and HLE-MM. Treat every one of those numbers as inclusionAI's claim until a neutral lab confirms them — but the direction of the tuning is clear and consistent.

The one number worth arguing about: 42 vs 38

The claim most likely to drive a text-only team to switch is the headline one: inclusionAI reports Ling-3.0-flash-VL scoring 42 on the Artificial Analysis Intelligence Index (v4.1.1), four points above the 38 it attributes to Ling 3.0 Flash on the same index version. Note what that index measures: its v4.1.1 composite draws on text and agentic suites — GDPval, Terminal-Bench, SciCode, GPQA Diamond, Humanity's Last Exam and similar, none of them image tasks. So the vendor is claiming that adding visual training to the shared backbone improved the model's text-side intelligence by four points, not merely that it can now describe pictures. That is a striking and entirely plausible claim — multimodal instruction tuning commonly lifts text ability.

Two sourcing caveats keep that number in proportion. First, the 38 is an earlier-index-generation figure: Artificial Analysis has since moved its live board to a newer index version (v4.3), where it currently lists Ling 3.0 Flash at an estimated 25 while still ranking it first among the 63 models in its class. The vendor's 42-vs-38 pair sits on the older scale, so it should not be read as "four points above where AA scores the text model today." Second, the 42 is inclusionAI's own figure on a model Artificial Analysis has not yet listed, and inclusionAI has not published the VL checkpoint's results on the individual text suites (SWE-Bench, Terminal-Bench) that its own text-model chart breaks out. Four points is exactly the size of gain you would want to reproduce before you migrate a text-only pipeline on the strength of it.

Price: one has a list price, the other doesn't

The cost comparison is currently a comparison with only one price in it. Ling 3.0 Flash is callable today on the vendor's first-party API at roughly $0.07 per 1M input and $0.22 per 1M output — vendor-set pricing, confirmed on Artificial Analysis's listing — with cached input cheaper, and its launch-period discounts have ended. Ling-3.0-flash-VL has no published API price anywhere; you cannot pay per token for it yet. If you want it this week, you self-host: MIT weights at roughly 250 GB in BF16, on the same hardware class the text model already wants — 4× 141GB GPUs (H20-3e or H200) or a 4-GPU Blackwell node at tensor-parallel 4, or 8× 80GB H100/H800 at TP8.

That reframes the economics for a text-only team. If you are buying tokens, the text model at $0.07/$0.22 is cheap and proven. If you are already self-hosting the 124B text model, the VL checkpoint is not a second infrastructure buy — it is another ~250 GB of weights on the same box class, justified only by workloads that actually consume pixels. The model with eyes only earns its keep when eyes are in the loop.

Who should pick which

• Text-only and high-volume — stay on Ling 3.0 Flash. You get a proven, AA-listed model, a real per-token price, and mature tool-calling. The VL model's four-point index gain is unreproduced and its text-side throughput is unmeasured; there is no evidence yet that it is the better text model, only a claim that it is.

• Vision enters the loop — documents, screenshots, charts, photos, video frames, or a UI your agent has to read and click — Ling-3.0-flash-VL is the obvious pick, because it is the same reasoning backbone you already understand with the vision stack added, rather than a separate multimodal model you have to re-evaluate from scratch.

• Mixed traffic, undecided — the low-risk move is to keep both reachable and route by request rather than re-architecting around one. That is the property a model gateway exists to give you: one API across 200+ models, provider list prices passed through at 0% markup, and automatic failover when a provider degrades. Neither Ling checkpoint is behind an OrcaRouter endpoint yet — the text model's price is the vendor's own and the VL model has no hosted price at all — but when the VL gains one, moving a subset of traffic to it and measuring is a configuration change, not an integration project.

What is still missing

• An independent run of Ling-3.0-flash-VL on a current Artificial Analysis Intelligence Index — the 42 is inclusionAI's word on an earlier index version.

• Any hosted API price for the VL model, which is the single biggest blocker to casual adoption.

• Published text-only splits (SWE-Bench Pro, Terminal-Bench 2.1) for the VL checkpoint, so a text-heavy team can verify that the vision stack did not cost it anything.

• An end-to-end latency or throughput figure for VL under image load — the text model's speed is measured; the vision model's is not.

• Neutral confirmation of the vision-benchmark chart, parts of which were run on inclusionAI's own harness and at least one in-house benchmark is slated for later release.

The verdict

This is not a model-quality contest; it is a modality decision with a four-point claim attached. If your inputs are text, keep Ling 3.0 Flash — it is cheaper, AA-listed, and measured, and the VL model's advantage over it is presently a vendor figure on an older index scale. If your workload starts from a screen — a document page, a screenshot, a chart, a video frame, a GUI your agent must operate — Ling-3.0-flash-VL is the same 124B brain you already understand, now able to see, and that is a genuinely new option that did not exist in the Ling fast tier a week ago. Adopt it where vision is real, validate the 42 before you migrate anything on the strength of it, and keep the choice reversible at the routing layer until an independent score and a list price arrive.