VRhino v0.9.4-alpha — Wan2.1 performance and memory

VRhino v0.9.4-alpha — Wan2.1 runtime and peak memory improvements

Install and downloads

curl -fsSL https://raw.githubusercontent.com/pixelrhino-ai/vrhino/v0.9.4-alpha/install.sh | sh

Linux x86_64, glibc 2.35+, a compatible NVIDIA GPU/driver and enough memory
and storage are required. The SHA-pinned installer downloads the core and its
exact private NVIDIA runtime. Existing verified R1 installations reuse that
runtime. No system CUDA Toolkit, cuDNN, FFmpeg or Python inference environment
is required; model weights remain separate.

Exactly two download assets:

• Core: vrhino-core-linux-x86_64-v0.9.4-alpha.tar.gz (36,211,068 bytes).
• NVIDIA Runtime: vrhino-nvidia-runtime-linux-x86_64-cuda12.8-cudnn9.8-r1.tar.gz (1,375,432,086 bytes).

The core contains new executables built as v0.9.4-alpha from
77bf96f5e6059f3643da378f274bbd58e7c5e835, based on merged performance source
55922514062f52fb2a9dc1f1cc03b725e9af443e.
This is not a metadata-only repackage of the old binary. NVIDIA R1, non-NVIDIA
libraries and media payload bytes are unchanged. Older releases remain immutable.

Main result: controlled VRhino before/after comparison

In the measured Wan2.1 T2V 1.3B / RTX 4090 D scene, complete generation time
fell 7.72% and sampled peak device memory fell 30.93%; checkpoints and video
outputs remained identical on the tested inputs.

Metric	Before five optimizations	After five optimizations
Full command wall time, median (range), s	306.356 (305.492–306.402)	282.720 (282.013–283.539)
Device-total peak, median (range), MiB, 1s sampling	13,941 (13,845–13,941)	9,629 (9,629–9,629)
Backend high water, GiB	13.25 (13.16–13.25)	9.04
Main-process HWM, median, GiB, 0.5s observation	11.355	9.413

Three complete fresh-process runs per side, interleaved, without a heavy
profiler or checkpoint export during timing. Wall time includes loading,
required verification, initialization, generation, media and process exit;
compilation/acquisition are excluded. Host HWM excludes media child processes.
Device samples are not strict instantaneous peak bounds; Backend statistics
are separate and are not added to device totals. File pages were already warm;
no system cache clearing, clock changes or GPU-resident cross-process service.

Same source base 8917863834cd64a59a3524e47b4425fabb8ffe12, with/without
only the five accepted changes. Optimized 5e5e0ea has the same tree as merged
5592251. This measured source A/B is not a direct benchmark of an older
published archive against this newly built archive, nor a sum of round percentages.
All six timing rows.

Task: vrhino/wan2.1-t2v-1.3b:1.0.1, prompt A red panda running through fresh snow,
seed5701, original negative conditioning, 832×480, 81 frames, 16 FPS,
50 steps, guidance5, flow shift5, second-order UniPC and unchanged BF16 policy.
Both source builds used Release/GCC11.4/nvcc12.8.93/SM89/cuDNN9.8/cuBLASLt12.8.4;
GPU RTX4090D/driver580.105.08, CPU Xeon Platinum8352V.
Resource budgets were unchanged: device21GiB, activation2GiB, host64GiB,
staging2GiB and safety1GiB.

The five general changes are request-owned RoPE position reuse, CUDA binary
broadcast indexing, runtime-dispatched AVX2 BLAKE2b with scalar fallback and
full-file verification, idle device resource reclamation after complete
sampling, and host-resident gather rows selected before upload. No model
identity special cases, task reduction, new precision or approximate algorithm.

Correctness and release validation

Two source A/B inputs (main seed5701; longer text/seed1234) each had all 413
available checkpoints, final latent/video tensors and all 81 decoded frames
byte-identical. Applicable generic/component/CUDA/lifecycle tests passed.
The new release build additionally passed 17 affected/image/version tests,
22 Distribution v2 tests, private dependency loading and isolated installation.
The packaged Wan2.1 normal Product smoke also passed using a sanitized PATH
and private dependencies: all 81 frames and the MP4 were byte-identical to the
validated 5592251 source output. Required Public CI gates the release PR.
This package smoke is an integration check, not another performance A/B.

The archive contains native images for SM80/86/89/90/120 plus compute80 PTX.
This binary is exercised on RTX4090D; image inclusion does not qualify every
GPU/model combination. Other real models, GPU families, CPU architectures,
Windows and macOS were not requalified for this release. Existing Windows
v0.8 assets remain unchanged. No smaller-memory GPU support is claimed.
The direct-Backend empty-batch indexed_gather SIGFPE remains a baseline issue
with an unspecified contract; it is not fixed here or treated as PASS.
Historical numerical HOLDs and existing gates/PrecisionPolicy are unchanged.

Additional official comparison — different execution strategies

Official source fixed at
9737cba9c1c3c4d04b33fcad41c111989865d315,
VRhino at 5592251, same machine and task/checkpoint provenance
Wan-AI/Wan2.1-T2V-1.3B@37ec512624d61f7aa208f7ea8140a131f93afc9a.
Solver/shift/50 integer timesteps match; sigma construction differs at most
5.96e-8. Prompt extension is disabled. Three normal fresh processes per side.

Metric	Official documented low-memory route	VRhino normal Product
Strategy/precision	T5 CPU BF16; DiT F32 parameters/BF16 autocast, F32 sampling state; GPU VAE F32; offload after sampling	GPU T5/DiT/VAE; existing BF16 policy, F32 statistics/residuals; idle resources reclaimed after sampling
Full wall median (range), s	425.318 (421.150–426.275)	283.307 (280.526–283.730)
Device peak, MiB, 1s sampling	13,129 each run	9,629 each run
Main-process HWM median (range), GiB	23.861 (23.782–23.862)	9.413 (9.412–9.413)

All six comparison rows.
Official environment: Python3.12.3, Torch2.8.0+cu128, FlashAttention2.8.3.post1,
cuDNN9.10.2.21, Transformers4.51.3, Diffusers0.39.0; VRhino uses cuDNN9.8 SDPA.
CPU OMP/MKL threads16. Official offload_model=True, t5_cpu=True unloads DiT
only after the entire sampling loop in this fixed source. Initialization
median86.596s and CPU conditioning41.734s explain much of the total difference;
remaining kernel-level attribution is unknown. Official allocated/reserved and
VRhino Backend metrics are not interchangeable. Main HWM excludes FFmpeg.
Official final finite-check overhead13–21ms is included in wall time, with no
large HOST export; all observed memory peaks preceded it.

Media defaults also differ: official FFmpeg7.0.2/libx264 CRF10, VRhino
FFmpeg4.4.2/libx264 CRF23. All six videos decode at the requested geometry/FPS,
with no blank frames or nonfinite raw output; sampled frames contain the
requested subject and snow. Composition/motion differ. Same seed does not
imply identical initial noise or video across frameworks.
No complete subjective video/temporal quality evaluation was performed;
there is no equal-quality superiority claim. The total difference reflects
these strategies/precision/defaults, not an unconditional Native-versus-PyTorch claim.

SHA256 and offline install

• Core: 3acd613cfc1b22bd0ca09d61fc5c68ce08408edcce984c92c1653a59df7815ac
• NVIDIA Runtime: c3488da9a4d346ff791cb14ba012018e50b7876ab7dd4dbecc7ba53e088dbe83

Offline: put both archives and the versioned tools/install_distribution_v2.sh
in one directory and run the tagged install.sh --bundle-dir DIR with isolated
--prefix/--bin-dir if needed. Checksums, exact runtime identity and helper
pin are in publication metadata. These records and
the helper are in tagged source, not additional Release download assets.

VRhino v0.9.3-alpha

VRhino v0.9.3-alpha

Install

curl -fsSL https://raw.githubusercontent.com/pixelrhino-ai/vrhino/v0.9.3-alpha/install.sh | sh

Downloads

• Core: vrhino-core-linux-x86_64-v0.9.3-alpha.tar.gz
• NVIDIA Runtime: vrhino-nvidia-runtime-linux-x86_64-cuda12.8-cudnn9.8-r1.tar.gz

The installer verifies both packages and manages the exact NVIDIA dependency
automatically. Normal users do not need to download the runtime manually.
First installation downloads both packages. A later core upgrade still
downloads the new core; if it requires this exact R1 runtime, the verified
private runtime is reused without another download.

Linux x86_64, glibc 2.35 or newer, a compatible NVIDIA GPU and driver, and
sufficient VRAM and disk space are required. No system CUDA Toolkit or cuDNN
installation is required.

This alpha organizes the distribution. It does not change inference or CUDA
numerics. The included executable was built as v0.9.1-alpha from source
commit df27ee5eb4b987fa4cb6e83d744f001d61883595; vrhino --version
reports that real binary build. The v0.9.3-alpha tag names this distribution
release. The historical Wan2.2 numerical HOLD is unchanged.

SHA256

• Core: 3f3ffc3fc1676c1616dac48749643136509ce299aec85e1d94479ccd9f26f4b0
• NVIDIA Runtime: c3488da9a4d346ff791cb14ba012018e50b7876ab7dd4dbecc7ba53e088dbe83

For offline installation, download both packages and the versioned
tools/install_distribution_v2.sh from this tag into one directory, then run
the versioned install.sh --bundle-dir DIR. The bootstrap checks the helper
SHA and both package SHA values. The existing repository
tools/rollback_distribution_v2_legacy.sh switches a migrated v0.9.1
installation back to its preserved copy.

The two package identities and installer helper pin are recorded in the
repository metadata. Neither this metadata nor
the helper is a Release download asset.

VRhino v0.9.2-alpha — Distribution v2

VRhino v0.9.2-alpha — Distribution v2

Public alpha prerelease; not stable/GA. This release promotes the exact v0.9.2-rc1 core and NVIDIA runtime archive bytes that passed public RC clean-host qualification. The archives were not rebuilt or recompressed.

Install

curl -fsSL https://raw.githubusercontent.com/pixelrhino-ai/vrhino/v0.9.2-alpha/install.sh | sh

The one-command installer retrieves a versioned, SHA-pinned catalog, then downloads and verifies exactly two artifacts: vrhino-core and vrhino-nvidia-runtime. NVIDIA userspace libraries are private to VRhino and selected by exact runtime identity. A compatible host NVIDIA Driver is required; system CUDA Toolkit, cuDNN, FFmpeg, Python inference environment, and sudo are not required. Model weights remain separate under vrhino pull.

The first install downloads both archives: 1,411,496,250 bytes total. Future core versions requiring the same exact NVIDIA runtime reuse the installed dependency and download only the new core. This core archive is 36,064,164 bytes. Existing installer-managed v0.9.1 monolithic installations are preserved for rollback; activation uses an atomic current pointer. Offline installation is supported with the same pinned assets.

Qualified asset identity

Asset	Bytes	SHA256
vrhino-core-linux-x86_64-v0.9.1-alpha-split-prototype.tar.gz	36,064,164	70ff9060357310b4c865790b92cb14749a91247534e282bbf8afbe97ca8594b0
vrhino-nvidia-runtime-nvidia-linux-x86_64-cuda12.8.90-cublas12.8.4.1-cudnn9.8.0.87-nvrtc12.8.93-nvjitlink12.8.93-r1.tar.gz	1,375,432,086	c3488da9a4d346ff791cb14ba012018e50b7876ab7dd4dbecc7ba53e088dbe83

Final catalog SHA256: d3890bd1b0d2d95f7be3500eb861abd84d2be4db4b77d70dfc600966d3a0fa80. Qualified Distribution v2 installer SHA256: b1780b4e1bc652b5a954a713497abf65b7f95665e814d48ca78e8e8c95b436e7. The versioned install.sh embeds both pins; the catalog .sha256 file is not its own trust root. Each archive contains internal checksums and provenance.

Qualification and provenance

RC evidence covers private NVIDIA loading, exact dependency reuse, offline install, legacy migration, rollback, and real Product paths: Wan2.1 PASS, Wan2.2 EXPECTED_HOLD_REPRODUCED, LTX PASS, Mochi PASS. Independent RC clean-host installation and Wan2.1 GPU Product smoke passed. A short external smoke from the formal v0.9.2-alpha URLs remains pending.

Wan2.2 runs and produces video through its normal Product path. Its historical generic numerical HOLD remains unchanged and is not claimed as a numerical PASS. Runtime, Backend, CUDA, PrecisionPolicy, and numerical gate semantics did not change.

The frozen binaries embed source commit df27ee5eb4b987fa4cb6e83d744f001d61883595 and v0.9.1-alpha build/version identity. The v0.9.2-alpha tag identifies this distribution publication commit. The historical v0.9.1 release and assets remain unchanged. Full URLs, installer instructions, rollback, and provenance are recorded in the versioned release record.

VRhino Distribution v2 Qualification RC v0.9.2-rc1

PRE-RELEASE — Distribution v2 Qualification RC

FOR CLEAN-HOST DISTRIBUTION QUALIFICATION. NOT A FINAL RELEASE. This RC stages the already qualified, frozen two-artifact Distribution v2 candidate for independent AutoDL testing. It is not a production installer migration decision.

The tag targets source/build commit df27ee5eb4b987fa4cb6e83d744f001d61883595, the same commit as v0.9.1-alpha-build-source. The core artifact retains its frozen v0.9.1-alpha-split-prototype identity; the v0.9.2-rc1 tag identifies this publication and does not claim a new Runtime build.

• Frozen core SHA256: 70ff9060357310b4c865790b92cb14749a91247534e282bbf8afbe97ca8594b0.
• Frozen NVIDIA runtime SHA256: c3488da9a4d346ff791cb14ba012018e50b7876ab7dd4dbecc7ba53e088dbe83.
• Runtime/Backend/CUDA/PrecisionPolicy semantics and numerical gates are unchanged.
• Wan2.2's existing numerical qualification HOLD is unchanged. Prior four-model Product comparison evidence is distribution-regression evidence, not a new Wan2.2 numerical PASS.
• A first install downloads both core and NVIDIA runtime. Later core versions can reuse the exact runtime identity.

vrhino-distribution-v2-catalog-v0.9.2-rc1.txt is the installer-readable, exact-ID catalog. The .json asset carries the same publication identity and asset sizes/hashes for independent inspection. The catalog SHA must be obtained out of band from the operator, not trusted merely because a .sha256 file is attached to this same release. The installer and legacy rollback helper are separate assets; neither is the current production monolithic installer. Verify the installer SHA from the operator before running it.

The independent test should cover anonymous HTTPS download, clean install, vrhino doctor, a real GPU Product smoke, same-runtime reuse, v0.9.1 monolithic handoff, and rollback. Do not treat this RC as stable/latest. If any asset needs changed bytes, use a new RC tag; never replace an asset under this identity.

VRhino v0.9.1-alpha — Linux CUDA

VRhino v0.9.1-alpha — Linux CUDA

Wan2.2 T2V A14B now uses the normal native Product commands, just like the other text-to-video models:

curl -fsSL https://raw.githubusercontent.com/pixelrhino-ai/vrhino/main/install.sh | sh
vrhino pull vrhino/wan2.2-t2v-a14b:1.0.0
vrhino run vrhino/wan2.2-t2v-a14b:1.0.0 --prompt "a red panda running through fresh snow" --output video.mp4

This Linux x86_64 package includes CUDA images for SM 80, 86, 89, 90 and 120 plus compute 80 PTX. It bundles the native converter and media helper; model weights are downloaded separately. A compatible NVIDIA driver, glibc 2.35 or newer and enough GPU/disk memory are required. Windows users can continue using v0.8.0-alpha.

Archive SHA256: 37bff0ecd0ac10d3bf2bd7c2c78f867477ca27a3c803d83f32a6d873ce015733 (1,412,112,885 bytes). The included SHA256SUMS verifies the extracted files. The installer checks both hashes before activation.

Representative native video runs have been completed on A800 and RTX 4090 systems using models that fit each device. This is an alpha release; generic numerical reference qualification continues separately. See the release record and Wan2.2 quick start for tested scope and setup.

Binary source commit: df27ee5eb4b987fa4cb6e83d744f001d61883595. The prior v0.9.0-alpha tag and assets remain unchanged.

VRhino v0.9.0-alpha — Linux CUDA

VRhino v0.9.0-alpha

This alpha brings Wan2.2-T2V-A14B native video generation to the Linux CUDA build,
along with schema2 package support and configurable weight-cache budgets.

Wan2.2

The native path includes text conditioning, sampling, VAE decoding and video
output. Three reviewed comparisons showed similar perceived visual quality and
prompt adherence to the official implementation.

On the tested A800 80 GB configuration (832×480, 81 frames, 40 steps, BF16), a
30 GiB weight-cache budget kept observed peak GPU memory around 44.8 GiB.
Generation took approximately 14–14.5 minutes. Performance varies by setup.

Wan2.2 currently uses vrhino evaluate. The archive includes a ready-to-edit
example under share/vrhino/examples/wan2.2/; weights are acquired separately.
Ordinary schema2 run support is pending. Existing schema1 entry points remain.

Download and start

Extract the Linux CUDA archive, then run bin/vrhino --version and
bin/vrhino doctor. An NVIDIA GPU and compatible driver are needed for inference;
Python and CUDA Toolkit are not required for native execution.
See the included example and documentation for model preparation and execution.

This release is Linux-only. Existing Windows users can continue using v0.8.0-alpha.

Alpha notes

Numerical alignment and wider configuration coverage continue to improve.
Start with the documented tested configuration when possible. A known F32
normalization edge case is documented in the developer notes; it was not observed
in the three reviewed BF16 videos.

Detailed qualification results are included in the developer notes and package
qualification records. Model weights are not bundled.

Source commit: 8584cec12016b18a822843362c357ada06e19cd8. Packaged binary embedded Git HEAD is unknown; the source archive and SHA256 manifest identify this build.

Installation and Wan2.2 setup

• Linux installation
• Wan2.2 native quick start: English · 中文
• Final release record and checksums

The current repository documentation supersedes preparation-status text in the frozen archives. Published assets and checksums are unchanged.

VRhino v0.8.0-alpha

VRhino v0.8.0-alpha is the first prerelease with a qualified native Windows x64 CUDA package. The frozen Linux and Windows binaries were built from d460de0a152c9ff92553b17ca37e4b5554c8c766; this release tag points to the later documentation-only commit b1d5ac2941353e6dbc82f8224ca41fb1da3c98f4.

Qualified assets

• Linux x86_64 CUDA: release readiness PASS; packaged remote import for all five public models PASS; Wan, LTX, MuseTalk, and LatentSync PASS.
• Windows x64 CUDA rc9: final clean-host qualification PASS on Windows 10 Pro 22H2 / build 19045, RTX 3090 24 GiB, NVIDIA driver 610.47; Wan, LTX, MuseTalk, and LatentSync native product paths PASS; packaged remote import for all five models PASS.
• Mochi remote import, install, doctor, and admission PASS on Windows. Full inference was not run on the 24 GiB qualification host. Its frozen minimum available device-memory admission remains 85,899,345,920 bytes. Historical Linux Mochi qualification is preserved.
• MuseTalk and LatentSync objective lip-sync quality gates remain deferred; no calibrated objective quality PASS is claimed.

The shared source-plan resolution fix is included. Source acquisition resolution changed; architecture, Runtime, backend, CUDA, kernel, precision, model-specification, and preset semantics are unchanged.

No model weights or converted VRMs are shipped. v0.7.0-alpha remains immutable and Linux-scoped. macOS support is not claimed.

See installation, the release record, and alpha limitations.

VRhino v0.7.0-alpha

VRhino v0.7.0-alpha

Alpha release for Linux x86-64 with CUDA. This release publishes the qualified native distribution and public production source; it does not expand the supported model set.

What's new

Full native production source and offline-capable developer builds are available. This release focuses on Runtime maturation, correctness, and repository sustainability; it does not add a new supported model.

Runtime architecture

LTX and Wan production self-attention use the bounded, typed, static NeuralGraph executor. Generic F32/BF16 values, explicit Cast, LayerNorm, Permute, Slice, Concat, indexed Gather, RoPE, Attention and SiLU compose through the shared native runtime. Preconditioned symmetric INT8 is a qualified foundation, not a claim that every model ships quantized. F32 numerical criteria remain finite and abs(err) <= 2e-5 + 2e-5*abs(ref). No model-specific backend or precision policy is introduced.

Correctness

VRM verification retains stable open-file identity. Component execution/tiling type ownership is corrected. Existing supported-model and generic runtime paths receive synthetic, external-reference and focused sanitizer regression checks.

Source availability

Public main is the only production source of truth. Project-owned source is Apache-2.0; upstream notices/licenses remain in place. Historical binaries and model weights keep their separate terms. Private research, media, weights, evidence, and build outputs remain external to source Git.

Build and packaging

Tokenizer sources are pinned and vendored for offline builds. Published public CI validates host builds without a GPU or model weights; CUDA qualification remains separate. The distribution includes CUDA 12.8/cuDNN 9.8 user-space libraries and bundled vrhino-ffmpeg. End users need a compatible NVIDIA driver, not Python, PyTorch, Diffusers, Transformers, Conda, CUDA Toolkit, system cuDNN/FFmpeg, GCC or CMake.

Supported models

Supported model package identities:

• vrhino/ltx-video-v0.9.1:1.1.1
• vrhino/wan2.1-t2v-1.3b:1.0.1
• vrhino/mochi-1-preview:1.0.1
• vrhino/musetalk-v1.5:1.0.1
• vrhino/latentsync-1.6:1.0.1

HunyuanVideo is not supported by this release. CogVideoX remains a canary and is not promoted to full release support. H3 has no production or Public support.

Known alpha limitations

Linux x86-64, glibc 2.35+, CUDA only; tested on Ubuntu 22.04/RTX 4090 D. An exact minimum driver has not been qualified. GPU/model memory requirements vary; Mochi's qualified default requires 80 GiB available device memory. Model pull/conversion can require substantial network and disk space. Native API v1 is local-first, has no authentication/TLS, uses serialized execution and in-memory job state. No Windows/macOS, general workflow scripting, universal architecture coverage, or stable C++ ABI is promised.

Upgrade notes

Keep the whole package directory intact. Native API v1, ProductInputSchema v1, model identifiers, and existing VRM v0.1 remain supported. Newly quantized metadata need not work with older runtimes. Rebuild C++ source integrations because internal construction/ownership types changed. Current project source licensing does not retroactively relicense v0.5/v0.6 binaries or third-party/model materials.

Public branding correction: the historical public banner and both README references are restored, and hygiene rejects the root banner and private asset directories. This documentation/branding/hygiene correction changes no runtime, model or package semantics.

Download and integrity

Release assets are vrhino-linux-x86_64-cuda-v0.7.0-alpha-final-candidate-5a07d09.tar.gz and vrhino-linux-x86_64-cuda-v0.7.0-alpha-final-candidate-5a07d09.tar.gz.sha256. Keep the distribution directory intact after extraction. The filename is retained from the exact qualified candidate; the archive has not been rebuilt or recompressed for publication.

Archive size: 1447181270 bytes. SHA256:

c0ae5a03ea5740535db707ac4bda8a996d83c03fceef61bdf23cda1fc1d3634e

Verify the download with:

sha256sum --check vrhino-linux-x86_64-cuda-v0.7.0-alpha-final-candidate-5a07d09.tar.gz.sha256

The release tag points to 39ae42cb9718af0dc4d43c92965ba88413c358c4. The distribution retains its recorded build identity 5a07d091a8fd5f86f45a412ec8e67e4415450233. The only intervening changes restore public branding and narrow its source-hygiene boundary; runtime, models, build/package inputs and qualification are unchanged. Embedded build-time provenance is preserved with the qualified bytes.

VRhino v0.6.0-alpha

VRhino v0.6.0-alpha

This feature Alpha adds machine-readable Product contracts and a built-in
local Native API for third-party and community integrations.

Highlights

• Inspect typed Product inputs, defaults, validation, and frozen Product facts
  with vrhino info MODEL --json.
• Start the local Native API with vrhino serve.
• Submit asynchronous Jobs, inspect status, cancel individual Jobs, and stream
  bounded NDJSON progress events.
• Use five immutable successor model packages.
• Benefit from substantial qualified LatentSync execution-time and memory-use
  reductions since v0.5.0-alpha.

Native API v1

vrhino serve starts on 127.0.0.1:11435 by default. Native API v1 supports
model discovery, Product schema inspection, run submission and status,
per-Job cancellation, and bounded NDJSON event streaming. See the
Native API v1 contract.

Native API v1 alpha is local-first. It has no authentication, TLS, or CORS and
is not intended for direct exposure to the untrusted Internet. Non-loopback
binding is explicit and emits a warning.

Machine-readable Product contracts

ProductInputSchema v1
declares required inputs, parameters and defaults, validation, outputs, and
frozen Product facts. Clients can inspect it with
vrhino info MODEL --json without parsing human CLI output.

Models

• vrhino/ltx-video-v0.9.1:1.1.1
• vrhino/wan2.1-t2v-1.3b:1.0.1
• vrhino/mochi-1-preview:1.0.1
• vrhino/musetalk-v1.5:1.0.1
• vrhino/latentsync-1.6:1.0.1

Pixel Rhino ships no original model weights or converted VRMs. Model and
content licenses remain independent from the VRhino binary license.

Performance and runtime improvements

Post-v0.5 work substantially reduced LatentSync execution time and memory use
through source-frame demand planning, optimized mask filtering, generic BF16
policy, exact Lanczos planning, overlapped verification, sampling-state reuse,
and SHA-NI verification acceleration.

Packaging and build

The Linux x86_64 CUDA package is self-contained above the compatible NVIDIA
Driver boundary. It requires no Python, Node, system FFmpeg, CUDA Toolkit, or
cuDNN installation. The production build was qualified with repository-local
pinned tokenizer dependencies and zero dependency-network acquisition during
the clean Release build and package process.

Known alpha limitations

Native API media inputs and explicit outputs use server-local absolute paths.
This alpha has no upload API, browser-direct CORS policy, authentication, TLS,
pull API, persistent scheduler, SDK, OpenAPI description, or official WebUI.
Product execution uses one worker; Job and retained event state are
process-memory only, restart-local, and bounded. See
Alpha limitations.

Download and verify

• vrhino-linux-x86_64-cuda-v0.6.0-alpha.tar.gz
• vrhino-linux-x86_64-cuda-v0.6.0-alpha.tar.gz.sha256
• Archive SHA256:
  9ebeabd2741850bc6811309bbc14683d613cfdf1d4c644b2c27012d9ea3db753

Download both assets, then run:

sha256sum -c vrhino-linux-x86_64-cuda-v0.6.0-alpha.tar.gz.sha256

See the installation guide.

VRhino v0.5.0-alpha

VRhino v0.5.0-alpha adds Public support for LatentSync 1.6 through Mode C local upstream acquisition and Native conversion. Pixel Rhino does not distribute LatentSync model weights or converted VRMs.

Highlights:

• bounded Native lip-sync diffusion from video and driving audio;
• deterministic --seed support;
• Public-safe BlazeFace + DWPose alignment;
• exact upstream verification and local Native conversion; and
• existing LTX, Wan, Mochi, and MuseTalk support remains available.

LatentSync model use remains subject to the CreativeML Open RAIL++-M License and its use-based restrictions. Users are responsible for lawful and consented input media. No upstream endorsement is implied.

Qualified on an NVIDIA GeForce RTX 4090 D; observed peak internal allocation was approximately 19.9 GiB for the qualification profile. This is an observation, not a hard minimum or admission threshold.