A flat vector hero on a charcoal background headed WeMM-Video-35B-A3B vs Runway Gen-4.5 with the subtitle two missing rows: a vertical column of slate rank bars of varying width with small square markers beside them, two of the positions left as empty dashed outline slots showing the background through, and a dotted divider.
Engineering & Research

WeMM-Video-35B-A3B vs Runway Gen-4.5: Two Missing Rows

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On the Artificial Analysis text-to-video board, read on 20 November 2026, there are thirty-five rows and neither Runway Gen-4.5 nor WeMM-Video-35B-A3B is one of them. The two absences are not the same absence. Gen-4.5 was placed at the top of that board on 1 December 2025 — 1,247 Elo, ahead of Veo 3 at 1,226, Kling 2.5 at 1,225 and Sora 2 Pro at 1,206 — in a launch post Runway still publishes. WeMM-Video-35B-A3B has never been on a board at all: Apache-2.0 weights committed to Hugging Face at 03:57 UTC that same morning, a technical report submitted the day before, and nothing from T​encent that resembles an announcement.

So one side has a retired crown and no current row, and the other has no crown, no row and no third-party number of any kind. That is a strange pairing to build a comparison on, and the honest version of it is narrower than a ranking. What is a board entry actually worth once the field has moved past it? What does each side publish that a reader can check today? And which of the two returns a file you can put in front of a client?

A crown is not a row

The launch claim is specific, sourced and still live. Runway's own research page for Gen-4.5 states that the model was ranked first on Artificial Analysis's text-to-video benchmark at 1,247 Elo, naming the three it passed. That is a real citation of a third party, made by the vendor, at a moment when the comparison was fresh. It is also the last independent-looking number Runway has printed for this model.

It cannot be laid beside today's table without arithmetic nobody has published. Artificial Analysis re-anchors its Elo scales between board versions: the current AA-Video-T2V V2.1 board fixes Kling 3.0 1080p (Pro) at exactly 1,000 as its reference point, ranks everything else against it, and prints a 95% interval on every row. Read on 20 November 2026, the leader is Wan 3.0 at 1,176 over 8,243 votes, and the anchor sits 23rd. A figure of 1,247 from December 2025 was measured against a different anchor, a different prompt set and a different field; dropping it into this table would imply a precision the two readings do not share.

What can be checked is presence, and presence has a shape. Across all three of the arena's video boards on the same read — text-to-video, image-to-video and video editing — the only row carrying a Runway logo belongs to Aleph 2.0, the company's in-context editing model, 11th on the editing board at 1,007 Elo over 23,082 votes. Gen-4.5 is in Runway's catalogue and callable this afternoon; it is not in the arena's current listing. A model leaving a leaderboard removes a marketing artefact, not a capability — and a model never entering one never had the artefact to lose.

A tall headless screenshot of the Artificial Analysis AA-Video-T2V V2.1 leaderboard as read on 20 November 2026, listing thirty-five ranked text-to-video models with Elo scores, 95% intervals, sample counts and per-minute prices, from Wan 3.0 at 1,176 over 8,243 samples down to Boreal at 827, with Kling 3.0 1080p (Pro) fixed at 1,000 as the board anchor, Veo 3.1 at 963 and $24.00 a minute in 26th place, and no Runway row anywhere on the list.

What the catalogue does say

Runway's developer documentation is more forthcoming than its research page. The identifier gen4.5 has been on Runway Dev since 10 February 2026, generates two to ten seconds per request from text or from a still, and bills at 12 credits per second — $0.12 at the portal rate of one cent a credit. That is a flat per-second rate, which makes a shot budget a multiplication rather than a simulation, and it is the single most useful thing on the page.

Delivery is where the ladder reveals itself. ProRes and PNG-sequence output add 5 credits per second to the 12. Ten-bit and deeper profiles, including half-float OpenEXR frames, add 20, rising to 40 once the output passes roughly four megapixels. HDR10, HLG, 12-bit PQ masters and bit-exact 16-bit PNG sequences have been available for gen4.5 since 20 August 2026, graded into BT.2020 with measured HDR10 metadata rather than tone-mapped afterwards. None of that is generation resolution — it is what the render is packaged into on the way out.

• Status — Gen-4.5: commercial product since 1 December 2025, API since 10 February 2026, no current arena row. WeMM-Video-35B-A3B: weights on the Hub, zero downloads, no announcement, never on a board.

• Direction — Gen-4.5: text or an image in, MP4, ProRes, PNG sequence or EXR out. WeMM: frames plus soundtrack in, text out.

• Shape — Gen-4.5: two to ten seconds, one generation per request, priced per second of output. WeMM: an 8,192-token joint context, sixteen frames sampled per clip, priced in gigabytes of resident weights.

• Price — Gen-4.5: 12 credits, $0.12, per generated second, plus 5 for ProRes or PNG and 20 to 40 for deeper colour. WeMM: 67.39 GiB of BF16, or 24.11 GiB in NVFP4, and the GPUs to hold it.

The ceiling written in shell flags

Divergence starts in where each one truncates, but it ends in whether you can find the number. Runway has printed its limits for months: two to ten seconds, four output ladders, a rate per second. T​encent printed a configuration and shipped commands that contradict it. The config says 262,144 positions; the invocations all say --max-model-len 8192, in vLLM and in SGLang alike. Between 8,192 and 262,144 there is no supported middle. The long number is headroom, and the serving flags are the product.

Convert and it stops being abstract. Ten audio tokens a second — a Whisper-shaped stack, 32 encoder and 32 decoder layers at width 1,280, then a window-level Q-Former at width 4,104 — puts about thirteen minutes of soundtrack inside 8,192 tokens, and that is before a single frame, the instruction or the reply is counted.

Picture budget goes next. The processor's defaults are 2 fps and up to 128 frames; the serving recipe replaces both with 16 frames inside a 262,144-pixel budget. On a ten-second clip that works out under two frames a second: a faithful transcript of the room and a sketch of what the room looked like. Which frames the report's own tables used is not stated anywhere, so the distance between its scores and your configuration is a number you cannot compute.

What the checkpoint lets you change

Here the two diverge in kind. Runway's output ladders, its 12-credit base rate and its two-to-ten-second window are product decisions; you buy the package or you do not. WeMM's 8,192-token window is a server flag, its 16-frame recipe is a line in a shell command, and its quantisation is a choice between 67.39 GiB of BF16, 22.69 GiB of AWQ W4A16 and 24.11 GiB of NVFP4. All three are yours to move, on hardware you own, at whatever latency you can afford.

The one thing the download does not give you is a supported engine. T​encent did not merge the architecture upstream; it shipped the diff instead. The patch directory carries an overlay per supported release — vLLM 0.18.0, vLLM 0.31.0, SGLang 0.5.16, SGLang 0.5.19, and no others — with a SHA-256 manifest beside each, and the compatibility surface under it is narrow enough to be brittle: one CUDA 12.9 toolchain, one PyTorch build, one Transformers build, Python 3.12, FFmpeg on the path, and a first SGLang launch that compiles kernels before it answers anything. Bring a version that is not on the list and the port is yours. T​encent's reference run spread BF16 across two H20s; the NVFP4 build fits a single 48 GB card at --max-num-seqs 1.

The weight is an artefact of the job. Hugging Face counts 36,177,564,536 parameters in the BF16 shards, of which the routed experts take the overwhelming share — 40 layers times 256 experts at intermediate width 512, top-8 selection, plus a 512-wide shared expert every token passes through — while about 1.0B are touched per token. Nearly 36B of capacity held in memory, 3B of arithmetic spent. Reading video instead of rendering it means paying to store knowledge the forward pass rarely visits.

A headless screenshot of the Hugging Face file listing for Tencent/WeMM-Video-35BA3B showing the Image-Text-to-Text tag, English and Chinese language tags, the arxiv:2611.26409 and apache-2.0 badges, the wemm_qwen3_5_moe architecture, LICENSE and SHA256SUMS files, a repository history of four commits, and BF16 safetensors shards from model-00001-of-00017 at 4.33 GB onward, uploaded about four hours earlier.

Transcripts have no export preset

The deliverable difference should decide a purchase before any benchmark does. Runway sells a finished file in an editable intermediate: ProRes for the grade, PNG or EXR for the comp, HDR metadata attached, an editor around it. WeMM-Video-35B-A3B sells nothing of the kind. What comes out is a paragraph — an account of what happened, who spoke, what was said — which goes into an index, a moderation queue, a QA report or a search box, not into a timeline.

T​encent's report claims that paragraph is the best available from open weights. Against Qwen3.5-Omni-Flash it scores 86.4 to 79.3 on Video-MME with audio and 54.9 to 36.6 on Video-MME v2 with audio; against its own backbone, Qwen3.6-35B-A3B, it wins 20 of 31 silent-video benchmarks. It loses too — JointAVBench 65.9 against 69.0, MUSIC-AVQA 79.5 against 82.1, EgoSchema 74.2 against 77.2. Every figure is the trainer's, the report says as much, and it switches between Instruct and Thinking mode table by table as it discloses. No third party has reproduced any of it. What would make the claim checkable is a vLLM upstream merge to end the patching and a first independent run on Video-MME v2 with audio.

The report also states that the family is "deployed across multiple core video services within WeChat," naming live-stream commerce, short-form drama analysis and interactive video. That is the strongest adoption claim in this piece and the least verifiable: no latency, throughput, concurrency or memory figure appears anywhere in the release. It is a sentence you can quote and cannot test.

A stacked flat vector evidence card on a charcoal background headed What each side has on the record: the bone-white Runway Gen-4.5 panel reads no row on any of the three AA video boards, 1,247 Elo at launch on 1 December 2025, 1,000 as the board anchor has since moved, and an API since 10 February 2026; the muted teal WeMM-Video-35B-A3B panel reads no entry on Artificial Analysis or LMArena, benchmarks reported by Tencent only, no third-party reproduction, and weights committed 20 November 2026. A grey footer reads 'Boards read 20 November 2026; the two absences are not the same absence.'

Where a transcript fits a studio

Put the two together and the pipeline writes itself, even though nobody has published it: render the shot, then have something read the render back — did the product appear, did the dialogue match the script, is this clip safe to publish. Runway does not sell the second half. A self-hosted WeMM checkpoint could, for the price of a 67 GiB download and a source overlay.

For the half you can have this week, the backbone is the sensible entry point. Qwen3.6-35B-A3B is the model WeMM-Video-35B-A3B was built on, as the report states, and it is on OrcaRouter at $0.25 per million input tokens and $1.48 per million output over a 262,144-token context with text, image and video input. It has no audio stack, so it cannot tell you who spoke. It can tell you whether the frame contains what the prompt asked for, against the same clips you render, today.

OrcaRouter serves neither Gen-4.5 nor WeMM-Video-35B-A3B, and the body of this piece would be a lie if it said otherwise. What the platform gives you is the harness for the comparison: one OpenAI-compatible key, provider list price passed through at 0% markup, and automatic failover when an endpoint degrades. Note also that Gen-4.5 bills in credits against a subscription rather than in tokens, which is not a unit a router can normalise — the credit is the meter, and the meter is Runway's.

Pick by deliverable, not by rank

Take Runway Gen-4.5 when the output has to leave the building as a file: a two-to-ten-second shot, delivered in ProRes, PNG sequence or HDR with metadata attached, inside an editor somebody else maintains. Go in knowing that the number on its launch page is almost a year old and was measured on a board that no longer lists it, that the 20-to-40-credit colour ladders add up quickly, and that the rate is flat because the package is fixed.

Take WeMM-Video-35B-A3B when the output is an account of footage rather than footage, when Apache 2.0 matters to legal review, and when you would rather move a ceiling with a flag than file a feature request. Go in knowing the weights are days old, that every quality number attached to them belongs to the team that trained the model, that three of the four quantised variants describe themselves as unvalidated, and that the install is a patch overlay on an inference engine that has not merged the architecture.

Two rows are missing from the board for opposite reasons, and neither absence is a verdict. One model was measured, was crowned, and has since been re-anchored out of the listing; the other has never been measured by anyone who did not train it. A comparison built on that has to be about what each side lets you do and what each one hands back — because the scoreboard, for once, is genuinely empty on both sides.