
GigaChat3.5-Omni Leak: Sber's 432B Model Grew Eyes and Ears in a Draft PR
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 185 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 228 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 98 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 199 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 113 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
A draft pull request opened in the SGLang repository on 18 November 2026 describes a model that does not exist in public: GigaChat3.5-Omni-432B-A28B, which is GigaChat 3.5 Ultra — Sber's 432B-parameter mixture-of-experts language model — with an InternViT-300M vision tower and a 24-layer conformer audio encoder attached, taking images, video with or without its soundtrack, audio files, and mixes of all three. Nothing in that sentence can be checked against a model card, because there is no public model card. The Hugging Face repository the PR links returns HTTP 401 when queried today, a name search on Hugging Face returns nothing, and Sber's own developer documentation still lists GigaChat 3 Ultra as its newest generation model, with no multimodal entry anywhere in the list. What exists is 3,290 added lines of serving code, a configuration file, roughly 2,000 lines of tests, and a benchmark table whose reference column somebody left blank.
That is not nothing. A pull request of this shape can only be written by someone who has had the checkpoint in front of them: the processor defers image decoding, pixel layout and log-mel extraction to the checkpoint's own remote code via --trust-remote-code, which means the author had to run it to know what those tensors look like. So the useful question is not "did Sber announce this" — it did not — but what the code says about a model that is very likely in a private repository right now, and what that would mean for anyone who currently pays per token to get audio and video understood.
What the pull request actually describes
SGLang PR #51276 is titled "[Feat] gigachat35-omni: Add GigaChat3.5-Omni image, video and audio support". It is a draft, it is unmerged, and its CI has been skipped precisely because it is a draft. The implementation is a thin wrapper — GigaChat35OmniForConditionalGeneration around the existing GigaChat35ForCausalLM — with three new things around it: a vision path through InternVisionModel and a pixel-shuffle MLP projector, an in-tree conformer audio encoder, and a processor that lays out the placeholders. Everything below is the PR author's own description of the model, and none of it has been confirmed by Sber.
• Language stack — unchanged from GigaChat 3.5: 432B total parameters, 28B active, 256 routed experts with 8 per token, 40 layers, hybrid attention mixing MLA with Gated DeltaNet linear-attention layers, two MTP heads. Those figures come from the public GigaChat 3.5 Ultra model card, so this part is verifiable.
• Vision — InternViT-300M, images split into 448×448 crops, 256 tokens per crop.
• Audio — a 24-layer conformer, 1,024 wide, 16 heads, 4,096-wide feed-forward, 64 mel bins, subsampling factor 4, attention restricted to 200-frame chunks with a halo of one block per layer. Audio tokenises at 6.25 tokens per second, one token per 160 ms.
• Video — the first frame keeps all 256 of its tokens; later frames keep 77 on average, and which 77 is decided after encoding, by ranking each token's mean absolute difference from the same position in the previous frame. A stable sort over the whole clip picks them, so a frame may end up keeping more or fewer than 77.
• Video with sound — use_audio_in_video interleaves each frame's share of soundtrack tokens after it, with HH:MM:SS stamps between frames.
• Long audio — clips longer than 32 minutes are encoded in windows with a halo, which the PR says is bit-identical to a single pass. Its claim: five hours of audio on one node.
• Failure behaviour — malformed media returns a 400 instead of crashing the scheduler, and SGLANG_VLM_CACHE_SIZE_MB defaults to 2048 for this architecture, because the stock 100 MiB could not hold a long video's embeddings and every prefill chunk re-ran the encoders.
Draft means draft
It is worth being precise about how much of this is verifiable today, because a draft PR with skipped CI is a proposal, not a fact about a shipped runtime. The file gigachat35_omni.py does not exist on SGLang's main branch — fetching it there returns a 404 — and the multimodal supported-models page in main has no GigaChat entry. The PR has one commit, from an account whose only SGLang interaction is this PR, and the sole comment on it is a documentation-bot note asking for preview approval from a fork.
The provenance is the strongest indirect evidence in the whole story. The processor's whole media path is delegated to "the checkpoint's remote processor", and the author reports prompt-id parity against "the serving stack the model was trained and evaluated with" — an internal reference stack that is not named and not public. You cannot write either of those against a model you have not run. Someone outside Sber's walls would need the checkpoint; someone inside them would have the internal stack. Nothing in the PR says which, and this piece will not guess.

The one thing the code makes clear is that this is not a rename or a config flag. The model registers a new model_type (gigachat3_5_omni) with its own token ids — 128004 for images, 128005 for video frames, 128006 for audio, 128007 for soundtrack — a new processor, a new embedding result type that mm_schedule.py has to dispatch on next to the existing EVS path, and new handling for chunked prefill, because the number of tokens a frame keeps is only known after it has been encoded. That is the signature of a model trained with those preprocessing decisions baked in, not of a community fine-tune hypothetically bolted onto a text checkpoint.
The token math is the part worth reading
Skip the benchmark table for a moment. The arithmetic in the PR is the substance, because it is what tells you whether a 432B model that ingests long audio and video is a thing you could afford to run — and it is the part that will not change if Sber never says a word about this model.
Take an hour of audio. At 6.25 tokens per second, that is 22,500 input tokens, plus a timestamp every 60 seconds and a closing duration stamp — small change next to the audio itself. Take a 512-frame video. At 256 tokens for the first frame and 77 per frame after it, the clip comes to about 39,600 tokens before its soundtrack, which is roughly a third of a 128K context for half a minute of footage. Add the soundtrack and each frame carries extra tokens on top. This is the whole argument for the change-reduction scheme: without it, 512 frames at 256 tokens each would be 131,072 tokens, an entire 128K window for one clip.
What that means in money is easy to check against models you can call today. On OrcaRouter's catalogue the mixed-media listings currently price an hour of audio like this:
• Gemini 3.8 Flash — $0.75 per million input tokens, so an hour of audio is about 1.7 cents, and a 512-frame clip about 3 cents.
• Qwen3-VL 235B A22B Instruct — $0.40 per million input tokens, so roughly 0.9 cents an hour and 1.6 cents a clip.
• DeepSeek V4 Flash Vision Exp — $0.22 per million input tokens, but image-only input; it cannot take the audio at all.
Put a real workload behind it. Forty hours of recorded calls is 900,000 tokens: about 68 cents on the Google listing and 36 cents on the Qwen one, at provider list price. OrcaRouter passes provider rates through with no markup, so a vendor price cut on any of these is live on our side the same day — which is also the honest limit of what we can say here: GigaChat3.5-Omni-432B-A28B is not in our catalogue and we are not a way to call it. It is not callable anywhere. The point of the arithmetic is that when something like it does appear, the audio is not the expensive part; the tokens you cannot compress are.
That is also where the PR's 32-minute window and five-hour claim sit. Those numbers matter to whoever self-hosts the model on eight H100s, because they decide how much of a long recording has to be re-encoded when a request is split across prefill chunks. They do not matter to anyone paying per token, where an hour of audio is an hour of audio whichever encoder produced it.
Four benchmark numbers, one blank column
The PR reports four scores, run on eight H100s in FP8 at tensor-parallel 8, greedy decoding: MMLU-Pro 73.07, MMAU test-mini 78.50, MMMU validation 64.11, and Video-MME 80.74 on a 512-frame clip with its soundtrack. Read them with three caveats attached, all visible on the PR page itself.
The first is that the reference column is empty. The table has a column for the model's original serving stack and a column for this PR, and only the second is filled, so there is no baseline printed next to any of these numbers and no way to tell from the table whether the port matches or loses to the stack the model was trained on. The prose does make a parity claim — identical prompt ids on every request the reference could serve, matching first tokens on 19 of 19 — which implicitly tells you the reference could not serve one of the twenty fixed requests.
The second is the checklist. The item reading "Provide accuracy and speed benchmark results according to Test the accuracy" is unticked, and the speed table, which has rows for batch 1 and batch 16 with columns for base decode, MTP decode, speedup and accept length, is entirely empty. So there is no throughput number for this model at all, despite the two MTP heads in the language stack that the public GigaChat 3.5 card credits with up to 2.2× generation speed.
The third is context, and here the numbers get interesting rather than merely thin. MMLU-Pro at 73.07 sits about 2.6 points below the 75.68 that GigaChat 3.5 Ultra's own model card reports for the text model on the same benchmark name — which may mean the multimodal variant gave up some text quality, or that the two runs used different harnesses, and there is nothing published that would let you choose between those explanations. MMAU at 78.50 is higher than the 74.7 that GigaChat's own audio model card lists for Qwen3-Omni, and far above the 62.2 the same card reports for Sber's own GigaChat3.1-Audio-10B-A1.8B — a card measured on the full MMAU set rather than the test-mini subset, so treat that comparison as suggestive and not as a like-for-like. MMMU at 64.11 is a mid-pack image-understanding score. Video-MME at 80.74 for a 512-frame clip with audio is the one that would raise eyebrows if anyone independent reproduced it.

What it takes to serve, and what Sber has said
Deployment reality first. The FP8 text checkpoint behind GigaChat 3.5 Ultra is 437.9 GB across 39 files, so tensor-parallel 8 on H100s leaves roughly 55 GB of weights per card before KV cache — which is why the PR's test rig is eight H100s and why that is not going to change much for the Omni variant. The two towers are rounding error against 432B: InternViT-300M plus a 24-layer conformer 1,024 wide is well under a billion parameters, so the hardware requirement is essentially the language model's. Metadata, launch flags and the gigachat35 tool-call parser are already in SGLang main from the GigaChat 3.5 work merged in September, and the project's cookbook has had a GigaChat 3.5 page since then; what does not exist in main is any multimodal path.
Sber itself has published nothing that this piece can find. Its developer documentation lists GigaChat 3 Ultra as the newest generation model, alongside GigaChat 2 Max, Pro and Lite, plus embedding models — no 3.5, no Omni, no multimodal entry. Artificial Analysis has no page for GigaChat 3.5 Ultra or for the Omni variant at any slug we tried today, so there is no independent score for the base model to reason from either. The public GigaChat 3.5 Ultra card is real, is MIT-licensed, targets Russian and English, and reports a vendor benchmark set including TAU2-bench 68.71, SWE-bench Verified 42.6, Terminal-Bench 2 at 13.48 and LiveCodeBench v6 at 56.2 — all Sber's own figures, unreproduced, and all for the text model only.
The audio lineage is the part of this that reads as a plan rather than an accident. Sber shipped GigaChat3.1-Audio-10B-A1.8B in July 2026: a conformer speech encoder and a modality adapter feeding a 10B mixture-of-experts decoder, trained for temporal grounding on a purpose-built long-audio dataset, with timestamped descriptions and long-audio summarisation as its headline skills. That is exactly the capability the Omni PR's stamping scheme is built to preserve — a timestamp before each frame, a stamp every sixty seconds of audio, a closing duration — and the conformer dimensions in the PR's config are consistent with folding that stack into the 432B model instead of sitting beside it.
What would settle it
Four things, in ascending order of how much they would tell you. A public checkpoint at that repository ID would confirm the model exists and let anyone read the config, which is currently the entire factual basis for the tower dimensions. A merged, non-draft PR with CI green would tell you the port works on more than the author's twenty fixed requests. A speed table with numbers in it would tell you whether the two MTP heads survive the multimodal wrapper, which is the difference between a model you can serve interactively and one you cannot. An independent benchmark — anyone's, on the full sets rather than test-mini — would tell you whether 78.50 and 80.74 mean anything.
Until then the decision for a reader is simple, and it does not depend on this leak resolving. If your product needs audio or video understood today, GigaChat3.5-Omni-432B-A28B is not on the table and no announcement has put it there; the models that can do the work are already priced per token, and the arithmetic above says the bill is smaller than most teams assume. If you are planning to self-host a Russian-language multimodal stack on your own H100s, this PR is the earliest public evidence that such a checkpoint is being built, and its token accounting is worth reading now — the compression scheme, not the benchmark table, is the part that determines your memory budget.
![The GitHub pull request page for sgl-project/sglang PR #51276, titled '[Feat] gigachat35-omni: Add GigaChat3.5-Omni image, video and audio support', showing 'Mogreine wants to merge 1 commit into sgl-project:main from Mogreine:gigachat35-omni', the tab strip with Commit 1, Checks 133 and Files changed 26, and the top of the Motivation section describing GigaChat3.5-Omni-432B-A28B as the GigaChat3.5 language model at 432B total and 28B active parameters with hybrid MLA and Gated DeltaNet attention, an InternViT-300M image and video encoder and a 24-layer conformer audio encoder loaded from the checkpoint.](https://cms.orcarouter.ai/api/media/file/4-4177.png)
What we would watch next: whether the repository goes public, whether the PR stops being a draft, and whether Sber's developer documentation grows a multimodal entry before either of those happens. Any one of the three would move this from a what-we-know-so-far piece to a launch, and this article would need rewriting rather than updating.
