Ling-3.0-flash-VL vs InternLumina U2: Both Shipped, Different Silicon

Both of these models are now real, runnable and downloadable, which was not true a fortnight ago — and the difference between them has almost nothing to do with benchmarks. Ling-3.0-flash-VL, from Ant Group's inclusionAI lab, put BF16 and FP8 weights on Hugging Face between September 4 and 8, 2026, published SGLang and vLLM serving recipes alongside them, and picked up an independent score of 25 on the Artificial Analysis Intelligence Index v4.3. InternLumina U2, from Shanghai AI Laboratory's InternLM team, shipped inference code first and then released its Ascend-trained model weights to Hugging Face on September 10 — seven safetensors shards in an ascend/ subfolder, rounding the repository out at sixteen files.

The catch is what those two weight drops were built for. Ling-3.0-flash-VL's published path is the CUDA one everybody already has: an eight-accelerator node runs the FP8 build, and the serving stacks are the ones you are already running. InternLumina U2's published checkpoint is the Ascend one, and the README is explicit that you download whichever checkpoint matches how you plan to run inference — with the NVIDIA-trained checkpoint still listed as coming soon. Two models released in the same week, and only one of them runs on the hardware most teams are standing in front of.

That reframes the comparison usefully. This is no longer a released-versus-vaporware piece, and anyone still describing InternLumina U2 as weights-pending is working from a week-old README. It is a piece about two very different unified-visual designs that arrived on two different silicon ecosystems, and about which parts of each release are actually finished.

What each release now contains

Ling-3.0-flash-VL is a 124B-parameter sparse mixture-of-experts activating roughly 5.5B parameters per token, on a 42-layer hybrid trunk alternating Kimi Delta Attention with gated MLA blocks at a 5:1 ratio. An arbitrary-resolution vision encoder and a two-layer MLP projector feed that trunk, with VideoRoPE handling spatial and temporal position. It takes text, images and video and returns text. Both BF16 (64 shards) and FP8 (roughly 126 GB, vision tower deliberately held at higher precision than the expert weights) are public under the MIT license, and the release is complete enough that the independent benchmark community scored it — 25 on the Intelligence Index v4.3, ranked #2 of 64 in its size class against a class median of 8. What it does not publish is a per-token price for the vision model, and its API examples carry an -rc1 identifier that leaves open whether the served checkpoint matches the open weights.

InternLumina U2 is a 16B-parameter sparse mixture-of-experts with about 1B active parameters, built on an LLaDA-2.0 MoE diffusion language model backbone, covering six task families in one model: text QA, text-to-image generation, image understanding, image editing, video understanding, and 3D understanding. Inference code for all of those is public on main. The Ascend-trained checkpoint is now public on Hugging Face. Still outstanding, per the repository's own roadmap: the NVIDIA-trained checkpoint, the training code (a separate repository), and the technical report. The roadmap ticks four boxes and leaves two open — inference code, Ascend weights, and the Ascend training and inference stack done; training code and tech report not.

One practical detail sits between those two paragraphs. Running U2's vision paths requires the AToken tokenizer weights at atoken_inference/checkpoints/atoken-sod.pt, and the released driver expects a split checkpoint directory with the model assets kept together. The code is real and installs against Python 3.11, PyTorch 2.6.0 and, on the CUDA path, flash-attn 2.7.2. Ascend support lives on a separate ascend-npu-support branch and needs CANN plus a matching torch_npu build in place of the CUDA toolchain. None of that is hidden; all of it is documented. It is simply a longer road than pip install and a model string.

Two answers to the question "what is a visual token"

The architectures disagree somewhere deeper than parameter count, and the disagreement is the most interesting thing about putting these two side by side.

Ling-3.0-flash-VL keeps the standard contract. An image becomes a sequence of tokens through the vision encoder and projector, those tokens enter a transformer trunk, and everything runs autoregressively. The novelty is not in how vision is represented but in how cheaply the trunk runs — 5.5B active parameters per token out of 124B total, which is the entire reason the model appears on the intelligence-versus-active-parameter frontier. VideoRoPE gives spatial and temporal position one consistent encoding so video needs no separate machinery.

InternLumina U2 changes the representation itself. It uses Apple's AToken tokenizer, in which each visual position is described by eight complementary codebooks — local texture, embedded text, geometry and high-frequency detail as separate streams rather than one collapsed vector. Each codebook keeps its own embedding on the way in, and the eight per-position embeddings are concatenated and projected down into a single backbone token. On the way out, prediction is what the team calls spatial-parallel, codebook-depth autoregressive: the diffusion backbone denoises many masked spatial positions at once, then a multi-codebook autoregressive head predicts the eight codes for each position in order. Understanding and generation run through the same machinery — which is the actual claim. The team's argument is that a single codebook caps how much information one visual token can carry, while discrete-continuous hybrids split understanding and generation across different representations and decoding paths, and that going fully discrete with multiple codebooks is how you get a genuinely unified model rather than two stacks bolted together.

Both of those are coherent positions, and they carry opposite costs. Multi-codebook representations can hold finer visual detail; they also multiply the token budget per image by the number of codebooks and make decoding considerably more involved, which is presumably part of why 16B-A1B was the scale chosen. Ling's sparsity buys cheap per-token inference; it also means a smaller active capacity per forward pass, which is the trade the class median of 8 on the intelligence index quietly illustrates — Ling-3.0-flash-VL clears it comfortably at 25, but it is not competing with dense frontier models on raw reasoning.

The task surfaces only partly overlap

Filing both under "multimodal model" oversells the overlap. Knowing where it ends prevents a lot of bad comparison shopping.

• Inputs — Ling-3.0-flash-VL takes text, images and video, up to 40 images per request, and returns text; InternLumina U2 takes text, images, video and 3D assets, and returns text, generated images or edited imagesbr /> • Image generation — Ling-3.0-flash-VL cannot generate or edit images at all; InternLumina U2's central claim is that it does both, up to 1024×1024, from the same weights that read imagesbr /> • 3D — Ling-3.0-flash-VL has no 3D path; InternLumina U2 ships a Blender-based pipeline that renders GLB and GLTF assets into 64 paired colour-and-depth views for the model to reason overbr /> • Video — Ling-3.0-flash-VL handles video through VideoRoPE temporal encoding; InternLumina U2 samples 64 framesbr /> • Context and output — Ling-3.0-flash-VL measures a 262K-token context window against the 256K its model card states, and a comparatively short maximum output; InternLumina U2 has published neither figurebr /> • Active parameters — Ling-3.0-flash-VL activates roughly 5.5B per token; InternLumina U2 activates about 1B, which is a fifth as much

Read that list and the conclusion is that these two are substitutes for exactly one job — reading an image and answering questions about it — and complements nearly everywhere else. If you need a model that generates or edits an image, Ling-3.0-flash-VL is not a candidate and no benchmark changes that. If you need one model that reads a chart, generates a variant, and reasons over a 3D asset, InternLumina U2 is aimed at that and Ling-3.0-flash-VL does not address it.

Benchmarks: one independent score against a page of vendor numbers

The evidence quality is not close, and it is worth being precise about why rather than asserting a winner.

Ling-3.0-flash-VL's 25 on the Intelligence Index v4.3 is a third-party run on a published protocol, with the class median at 8 for context, plus measured throughput of 142.9 output tokens per second against a peer median of 105.1 and a 2.21-second time to first token. Its vendor card separately claims 42 on the retired v4.1.1 protocol, which is a different scale and cannot be set beside the 25 as though the model declined; the index was restated in September 2026 and re-scored across the board. The vendor also reports a 1,441-versus-1,440 Image-to-WebDev Arena win over GPT-5.4 under the handle "linthium" — a one-point margin inside Elo noise, and vendor-reported.

InternLumina U2's numbers are the lab's own, labelled preliminary and partial, with the full comparison tables and technical report still pending. They now sit in a table on the repository README rather than on a benchmark leaderboard, and they cannot be independently verified yet because no third party has published a run. Stated as the lab states them:

• Understanding — ChartQA 86.52, CharXiv-DQ 83.65, HallusionBench 62.15, MMMU-Pro 36.13, MathVision 33.22, DynaMath 56.37br /> • Video — Video-MME 51.26, MVBench 59.74, MLVU 57.03br /> • Generation and editing — GenEval 0.81, DPG 87.10, TIIF Short 84.60, TIIF Long 85.95, ImgEdit 3.83br /> • Sourcing — every InternLumina U2 figure is vendor-reported and preliminary; every Ling-3.0-flash-VL figure is vendor-reported except the Intelligence Index, which is per Artificial Analysis

Two of those deserve a flag. GenEval 0.81 trails a named competitor's 0.89 by the lab's own table — a rare instance of a vendor publishing a loss, and a reason to trust the rest of the sheet slightly more. And ImgEdit at 3.83 is meaningless without the accompanying table, which is part of what the pending technical report will carry. Treat the InternLumina U2 list as results the lab intends to defend, not as results you can act on. Note also that its understanding scores and Ling-3.0-flash-VL's index score are not comparable quantities: one is a set of task-specific vendor numbers, the other a composite third-party index.

License, hardware, and what a bet on each actually costs

Ling-3.0-flash-VL is MIT across both precision formats, which is about as clean as open weights get — no riders, no field-of-use restrictions, no ambiguity about commercial deployment. The FP8 build at roughly 126 GB fits within a node of eight 80GB-class accelerators; BF16 is roughly twice that. Serving support already exists in SGLang and an Ant-maintained vLLM fork. The residual risk is commercial rather than legal: Ant publishes no per-token rate for the vision model, free access on Ling Studio is promotional rather than a tariff, and Artificial Analysis lists it at $0.00 per 1M tokens in and out only because the endpoint it can see is the free one.

InternLumina U2 is Apache-2.0 on the main repository, with two carve-outs worth reading: the bundled VeOmni/ directory retains its upstream Apache-2.0 license, and atoken_inference/ is derived from Apple's ml-atoken and redistributed under Apple's original terms. The infrastructure claim is serious — it targets both NVIDIA GPUs and Huawei Ascend NPUs with operator-level numerical alignment between the backends, and the team reports training end-to-end on a thousand-card Ascend cluster with a 1.85× training-efficiency improvement from fused operators, tuned HSDP sharding and gradient checkpointing. That is real engineering. It is also orthogonal to whether the model is any good: training efficiency on Ascend says nothing about output quality, and the checkpoint that exists today is the one aimed at that stack.

So the practical asymmetry is not open-versus-closed. Both are open, both are permissively licensed, and both can be downloaded today. It is that one release assumes the hardware you probably have and the other assumes the hardware you probably do not. If your fleet is NVIDIA, Ling-3.0-flash-VL is an afternoon's work and InternLumina U2 is a wait. If your fleet is Ascend — which is increasingly common in the Chinese market this model was built for — that equation inverts completely, and InternLumina U2 is the one with a finished path while Ling-3.0-flash-VL's recipes assume CUDA.

Questions people actually ask about this pairing

Are InternLumina U2's weights downloadable yet?

Yes, the Ascend-trained checkpoint is — seven safetensors shards published under ascend/ on Hugging Face, alongside the config, tokenizer and modeling files. The NVIDIA-trained checkpoint is a separate subfolder and is still listed as coming soon, and the training code and technical report remain outstanding.

Is Ling-3.0-flash-VL strictly better because it has an independent score?

No — it is better-evidenced and easier to run on common hardware, which is a different claim. Ling-3.0-flash-VL cannot generate or edit images, cannot process 3D assets, and has no published price. If your task is image generation, image editing or 3D understanding, InternLumina U2 addresses it and Ling-3.0-flash-VL does not address it at all, however many benchmarks the Ling model has.

Does the 42 on Ling-3.0-flash-VL's model card contradict the 25 from Artificial Analysis?

Not directly. The 42 was measured against the Intelligence Index protocol v4.1.1 and the 25 against v4.3; the index was restated in September 2026 and re-scored across the board, and the two versions are built from different evaluation sets. What can be said is that the 42 has never been independently reproduced and the 25 has been independently produced.

Where either of these can be called

Neither Ling-3.0-flash-VL nor InternLumina U2 is on our model catalogue, and saying otherwise would be a claim a reader could check in ten seconds. Ling-3.0-flash-VL is reachable through Ant's own platform, which publishes an OpenAI-compatible endpoint and an Anthropic-compatible one, through free trial access on Ling Studio, and through the open weights if you serve them yourself. InternLumina U2 is reachable through the Hugging Face checkpoint and the repository's inference driver, on the hardware that checkpoint targets.

What we do route is the vision model this pairing most often gets weighed against in practice: DeepSeek V4 Flash Vision Exp, live on the catalogue with a 1M-token context window and a published rate, on the same OpenAI-compatible endpoint as the rest of the catalogue. When a lab releases open weights, a routing layer can usually offer the model without waiting for the lab's own hosted service to stabilise — which is the practical difference between a model being quotable and a model being callable. That difference is live for Ling-3.0-flash-VL and, for the moment, academic for InternLumina U2, whose release story is a hardware path rather than a hosting question.

The verdict is genuinely split, and it splits on hardware and task rather than on quality. Ling-3.0-flash-VL is the complete release: MIT, both precision formats, third-party scored, CUDA-first, and limited to understanding. InternLumina U2 is the more ambitious design and the less finished release: one hardware stack's checkpoint published, a unified six-task surface including generation and 3D, and a technical report still owed. If you need a vision model on NVIDIA hardware this week, the choice makes itself. If InternLumina U2's multi-codebook design is what interests you, the thing to watch is the NVIDIA checkpoint — and the honest position until then is that its benchmark table, including the row it loses, describes a model whose second half has not shipped.