China's Lion-ER-35B-A3B tells "higher in the picture" from "higher in the room", and the weights are open

On 4 November the Shizishan lab (狮子山, "Lion Mountain") at the China Merchants Advanced Technology Institute open-sourced Lion-ER-35B-A3B, a vision-language model trained not to describe a photo but to work out what stands where in a room and how far along a task is. Thirty-five billion parameters in a mixture of experts, about three billion active per token. The size is not the interesting part. The team wrote down 39 specific skills first and only then went looking for data, instead of pouring in everything it could find, and on ten public embodied benchmarks all ten scores rose against the base model, with an equal-weight average of 65.29.

These models are not chat assistants that caption pictures. The idea is to give a robot eyes with a brain attached: frames from its camera plus a text task go in, answers like "still a long way from the handle", "the camera turned left" or "that crate is the same one from the previous frame" come out.

Three things a robot keeps getting wrong

The world is three-dimensional, the frame is flat. Most failures grow out of that one fact. Move the viewpoint and the same cabinet stops looking like itself. An object that shifted in the frame may not have moved at all — the camera did. And "approaching the sushi roll" and "reached the sushi roll" are a couple of centimetres apart on screen, which decides whether it is time to grab.

The clearest example is a bottle. On two consecutive frames from a robot arm's camera it slid upwards. Either the bottle was lifted, or the camera tilted down. The robot has to do opposite things in those two cases: reach higher or reach lower.

The release names exactly these three failure classes: confusion from a change of viewpoint, a misread camera motion, and a misjudged task stage. None of them is about recognising objects. All three are about space and time.

A list of skills first, data second

The obvious route is to hoard more robot datasets. The lab went the other way and says it plainly: more data is not more capability. It is a known result that even reweighting domains inside the same dataset shifts downstream quality noticeably.

So they drew a capability blueprint first. Eight top-level domains: entities, space, time, physics, affordance (what can actually be done with a thing), task, social context, evidence. Inside them, 39 second-level skills. For every sample they also record separately how the model is supposed to reason, what goes in, and what shape the answer takes.

Then comes "capability compilation": an abstract requirement is turned into a concrete task. How many viewpoints, what has to be visible, which question to ask, what counts as proof of the answer. If the frame does not carry enough information for an honest answer, the task never enters training. A guess is not a sample.

Answers are checked through three routes. Cleaning old data: frame order, what "this" refers to, whether the options and the key actually match. Geometry-built tasks, where camera pose and depth are reliable, so the answer is computed from the scene rather than invented. And targeted synthesis where observations are thin, with one rule: a generated answer is never ground truth on its own and is re-checked separately.

The final fine-tuning set is 312,380 samples. The training budget was allocated by which skill needed work, not by which dataset happened to contribute the most rows.

The second loop is an eval-driven data harness, where evaluation is the start of the next data round. Every model error becomes a request for new examples: from a pool of 31 sources and roughly 4.3 million records they pull material with the same reasoning logic in a different setting. One example from the release: the model kept treating "higher in the picture" as "higher in the room", so the next round added scenes with multi-level shelves, occlusions and perspective distortion.

What the benchmarks say

Everything was measured at avg@4 — each task answered four times, scores averaged. The comparison base is Qwen3.6-35B-A3B, the same size and architecture without the robot-specific training.

• Where2Place, where an object can be placed: 64.28, up 9.30 on the base.
• ERQA, general embodied reasoning: 64.38, up 6.19. The trajectory sub-category reaches 68.18, up 12.12.
• SAT, spatial relations: 89.33.
• Equal-weight average across ten benchmarks: 65.29, with all ten above the base.

The benchmark list is unusual: Where2Place, ERQA and SAT are narrow, young sets built around robot questions. Convenient for a release note, less so for a head-to-head with the flagships.

Four tasks where the base models wobble

The release walks through four examples from public sets, and they say more than the summary scores.

The first is that bottle. Two frames from an arm-mounted camera, the bottle slides up the image, and the question is which way the camera turned. The base model got it right once in four tries, Qwen3.6-27B never, Lion-ER four out of four: the camera tilted down.

The second is trajectory choice. Four coloured paths on screen, the job is to unplug a charger and carry it to the right. Both base models bit on a long purple arc that sweeps beautifully towards the goal but never touches the charger. Lion-ER ruled the options out one by one and picked orange: reach first, then carry.

The third gives three photos of one flat, with a butterfly painting visible only in the last. Where was it relative to the photographer in the first shot? Both base models said left. Lion-ER said behind, and that is correct — you have to rebuild the floor plan in your head.

The fourth is progress. A video shows a robot about to put a sushi roll into a pot, and the question is whether the step "move to the roll" is already complete. The base model said yes three times out of four, while the arm was still approaching.

Where the catch is

Every number comes from the lab's own release. There are no independent runs yet, and avg@4 is four attempts — thin for fine distinctions.

More to the point, the model beats its own base, not the frontier. The release contains no direct comparison with top closed models beyond one Chinese model from the same family in the worked examples.

And answering questions about a scene is not the same engineering job as driving an arm. Between "the model knows where the handle is" and "the robot took hold of it" sits a layer of planning, calibration and safety that the release does not cover. Three billion active parameters for long-horizon scenarios is modest too.

The good news is that the weights are out: Apache 2.0, BF16 Safetensors, running through Transformers, vLLM and SGLang. This is no longer "watch the demo", it is "take it and test it".

Our own numbers: how long an image takes and what it costs

While reading the release I ran the numbers on our own database. "The model understands space" without speed and price sounds like a press release rather than a conversation.

Over 30 days, from 6 October to 5 November 2026, neuralspace.pro ran 16,867 image generations. 831 of them failed — 4.9%. Here is the interesting part: 640 of those 831 failures, almost eight in ten, were not timeouts or outages but content filters firing. The model simply refused to draw what was described. When it feels like the service is broken, most of the time the service is fine and the filter is not.

On the cheap models over the same 30 days: z-image ran 187 completed generations with a median of 13 seconds, a 95th percentile of 215 seconds and an average cost of $0.0040 (about 0.35 ₽) per image. nano-banana: 695 generations, median 23 seconds, p95 162 seconds, $0.0200 (1.75 ₽). seedream-45: 1,119 generations, median 59 seconds, p95 169 seconds, $0.0325 (2.85 ₽). Eight times the price and four to five times the median — exactly the case where drafts and storyboards are cheaper on the smaller model.

What to do with it

The weights are on Hugging Face (cmriat/lion-ER-35B-A3B), and the card promises fine-tuning for your own scenes: a workshop floor, a warehouse, a lab bench. If you want to poke at models like this without a GPU of your own, there is the chat and the Video section; images are generated in Images.

Source: 量子位 (QbitAI), 4 November 2026 and the model card on Hugging Face.