Murea open-sources InWorld: a 42 GB model

Murea Labs published InWorld on October 17, 2026: an interactive world-model project that pairs a world-state ledger with an audiovisual generator. The Hugging Face repository MureaLabs/InWorld took a 42,018,192,440-byte checkpoint (about 42.02 GB, not gated) at 05:19Z that day — uploaded as model_ema_distilled8_step_07000.safetensors and renamed to murea-fast.safetensors at 05:23Z; the code repository merged PR #2 the same morning under Apache-2.0 and ships a 27-page technical report. The weights that exist are Murea-Fast, the 8-step distilled model. The 81.13 that tops the WBench Full-289 table belongs to Murea-Pro, whose weights were not released. This Murea Labs is unrelated to Inworld AI (inworld.ai), the realtime-voice and game-NPC company that search results tend to surface first.

Key points

The model card states the checkpoint’s size, sampling and output: 42,018,192,440 bytes, 8 steps, 1280×704 at 24 fps, with a 121-frame example clip of about 5 seconds.
The performance doc measures four 141 GB H20s at 16.1 s for a first 121-frame request and 15.4 s for a repeat, with 78,852–80,866 MiB reserved per GPU; the README rounds this to “about 17 s”.
The WBench Full-289 split lists 16 models. Murea-Pro leads at 81.1, XGEN-Jing follows at 81.0, a gap of 0.1; the default Navi(158) split has 44 models and does not include Murea at all.

What actually shipped

InWorld is two pieces. Murea-Mind is an agent that plans, checks and remembers; Murea-Fast is the service that turns references and prompts into video with synchronized sound. The repository keeps them separate — inference/ holds the Fast service, agent/ holds the Murea-Mind app that opens a world and runs a story — and they talk over a documented HTTP API. That split explains the “world model” label: the generator shoots, and a separate state ledger keeps the cast and the place consistent between clips.

The weights are not a download-and-run artifact. The deployment guide says the checkpoint replaces the DiT; inference/README.md is more specific: it holds the DiT plus 258 trained text-connector tensors, and prepare merges those 258 tensors into the base components while keeping the four text-projection tensors that exist only in the base (audio_text_proj_in and video_text_proj_in, weight and bias). The checkpoint therefore cannot serve as a drop-in DiT replacement on its own. A full LTX-2.5-Diffusers tree is still required, with the tokenizer, text encoder, connectors, video and audio VAEs, and vocoder. The documented path is to verify the SHA-256, run prepare, then start, and poll 127.0.0.1:30260/health. The README’s security section says plainly that neither the inference service nor the agent has authentication, so a shared network needs a look at SECURITY.md first.

The speed numbers are more useful broken out by request type than as a headline. The four-GPU configuration is 141 GB H20s with sequence parallelism 4; the one-GPU column is the same machine, not a different optimization.

Request	One H20	Four H20s	Peak memory per GPU (four)
121-frame generation, first request	44.8 s	16.1 s	78,852 MiB
121-frame generation, repeat	43.7 s	15.4 s	79,766 MiB
137-frame continuation (17-frame prefix)	44.8 s	18.8 s	80,866 MiB

Two limits come with that table. References are not resized to the output resolution: a 768×768 image adds 576 tokens and a 4000×3000 photo adds 11,750, so latency climbs with reference size. The same document reports that with three characters, about half the clips at every tested size lost a character or merged two, and calls that case unreliable.

Building the top row

WBench is an interactive-video world-model benchmark from Meituan LongCat and Fudan University; its page says it was updated 2026/10/12, and Full-289 covers 289 cases and 1,058 interaction turns. On that page the Murea-Pro row carries data-model-id="museflow_pro" and an icon at imgs/musee.png, and its listed update date is 2026/09/23 — earlier than the repository’s October 9 commit labeled “Murea first public version”. The technical report’s PDF metadata still titles the document “Musee: Reference-Conditioned and Streaming Audiovisual Generation” and names the author “Musee Model Team”. The rename happened before publication; this is what the page and the file metadata show, not something Murea describes.

The shape of the table matters. Full-289 has 16 rows; Navi(158) has 44, and the page opens on Navi. Murea appears only in Full. Navi’s top three are Alaya-EVOKE-Turbo at 82.0, XGEN-Jing at 81.9 and Enuma at 81.7, all above Murea-Pro’s 81.1. In Full, second place is XGEN-Jing’s bidirection configuration at 81.0, third is its 4-step AR configuration at 79.9, and Kling 3.0 is fourth at 79.4.

Rank	Model	Overall	Status
1	Murea-Pro	81.1	Unreleased, self-evaluated
2	XGEN-Jing (bidirection)	81.0	Open source, self-evaluated
3	XGEN-Jing (4-step AR)	79.9	Open source, self-evaluated
4	Kling 3.0	79.4	API

The self-evaluation flag needs precise wording. In the page source the Murea-Pro row carries data-evaluation-source="self-evaluation", and nothing in the page consumes it: the scripts call getAttribute twice, on href and src, and no CSS selector matches the attribute. Strip script and style and the visible text contains “self” zero times. Murea is not singled out — the two XGEN-Jing rows, Alaya-EVOKE-Turbo and Enuma carry variants of the same attribute, and the page hides it for all of them. “The leaderboard marks the row as self-evaluated” is true only if you read the HTML.

Three places the framing slipped

The first is the model. The top row, Murea-Pro, is evaluated in section 5.4 of the report at 50 sampling steps, I2VA first-frame conditioning, 1344×768 at 24 fps and a fixed seed of 42. The released model, Murea-Fast, runs 8 steps at 1280×704. The abstract’s highest score and the released weights belong to different models; coverage that reads “first on two leaderboards” merges them. The leaderboard itself does label the row Unreleased, so that part was not hidden.

The second is the denominator. 81.13 is first out of 16 on the leaderboard, not first among everything evaluated — Table 3 of the report lists 14 rows and leaves out XGEN-Jing, Alaya-EVOKE and Enuma. The “leads the runner-up” claim is shakier still: the report picks Kling 3.0 as its highest baseline and computes +9.52 on Interaction and +5.23 on Physical. The actual runner-up on the leaderboard is XGEN-Jing, with 76.7 on Interaction and 73.3 on Physical — gaps of 5.9 and 1.1. The October 18 Machine Heart article reports the 9.52 and 5.23 figures as leads over “the runner-up”, which treats a chosen baseline as second place.

The third is the 64.98 on OpenS2V-Eval. Table 1’s caption says all scores were recomputed with a common evaluation pipeline, so this is not the number the benchmark publishes. Seedance 2.0 scores higher in that table, 67.13, but returned 148 clips out of 180 with 32 refusals; on the 148 shared cases Murea-Fast scores 66.11. On OS2V-Pro, Murea-Fast’s aggregate is 69.9 against Kling V3 Omni’s 70.1, and the report’s own prose calls it “close to”. Together these support “among the strongest open reference-to-video models”, not SOTA.

What the report concedes

Page 19 of the technical report contains a disclosure that is easy to miss: OS2V-Pro’s VTF and Bind checklists are derived from the expanded prompts given to Murea-Fast, and those checklists contain details absent from the official prompts. The report calls this “an input advantage” for Murea-Fast. The task-following score therefore mixes prompt information with model behavior. The report’s own component scores mark the boundary: VTF is 95.5, first, while Bind is 19.6, below Kling V3 Omni’s 22.6. Audio is weaker too, at ATQ 59.7 and VESA 75.0 against Kling’s 66.3 and 88.7.

Task-level scores are uneven in the same way. Murea-Pro scores 90.92 on event editing and 87.04 on subject action, 80.65 on perspective switching, and 71.79 on navigation. The conclusion names navigation, long-range planning and setting consistency as the main areas for improvement. Those are the authors’ words, not an outside critique.

License and first steps

The code is Apache-2.0; the weights are not. The model card marks the license as other and points to the LTX-2.x Community License, dated August 11, 2026, and the README says the checkpoint is a derivative of LTX-2.5. Section 2.1 requires any entity with annual revenue of at least $10,000,000 to obtain a paid commercial license for any use other than a purely non-commercial purpose. Calling this fully open source would mislead a team planning to ship it. Open weights and an open-source license are separate things here, the same distinction we drew for Ling-3.1-flash.

The shortest path to running it: hf download MureaLabs/InWorld murea-fast.safetensors --local-dir ./weights, clone https://github.com/Murea-Labs/InWorld.git, follow the deployment guide to fetch the LTX-2.5 base components and set four absolute paths in configs/, then run bash scripts/murea-fast verify <checkpoint> --sha256 followed by prepare and start. Murea-Mind runs on any machine with Python 3.12, but it needs an API key for a text and image model — Gemini, an OpenAI-compatible server, or DashScope.

Our take

This is worth a reader’s time because the 42 GB checkpoint, the memory and latency figures and the 27-page report are real, and the documented path actually runs: 15–16 seconds for a 5-second clip with sound on four H20s is workable for a team doing reference-to-video. We covered a different speed story in SparkDiffusion; that one is a single-GPU speedup, this one multi-GPU throughput, and the two numbers cannot be compared directly.

Carry three caveats. The leaderboard leader is a 50-step model nobody can download, while the release is an 8-step model, so the two should never be quoted as one result. The top of that table is a 0.1-point gap, 81.1 to 81.0, and the runner-up carries the same hidden self-evaluation flag. And Murea-Fast sits on LTX-2.5; how much of its quality comes from the base model is not ablated anywhere in the report. Against a closed commercial system such as Kling 4.0, InWorld’s advantage is that it can be downloaded and modified, and its cost is the license and the setup work.

The community has not shown up yet. On 2026-10-19 the GitHub repository had 6 stars and 2 forks, the Hugging Face model 0 downloads and 2 likes, and a Hacker News search by title and URL returned nothing; there is no third-party reproduction or independent measurement. The only Chinese coverage we found is one Machine Heart article; its original URL served a promo page when fetched directly (3,375 bytes titled “机器之心·数据服务”), so we read the text through a proxy and cross-checked it against a NetEase repost. Every judgment here comes from primary material, with no independent newsroom to lean on.

The next things to watch: whether Murea-Pro’s weights ship, and whether WBench ever renders the self-evaluation flag — right now the page gives a reader no visual way to tell an organizer evaluation from a team’s own.