Every build, dated.

cuda-wan-1.2.92 · veizik WAN (CUDA) v1.2.92 — Wan 2.1 & Wan 2.2 video Current

Text-to-video and image-to-video with Wan2.1-T2V-1.3B, Wan2.1-T2V-14B, Wan2.2-T2V-A14B and Wan2.2-I2V-A14B on your own NVIDIA GPU. Veizik runs the models through its own native engine: no Python, no CUDA toolkit. You supply the weights; the frames are written as PNG files to the folder you name. It also runs a resident render server with a host-loop (bin/veizik-resident) that keeps one process and one GPU context across requests, and mid-render preemption. Distributed as a single signed .vzpack, and render-verified fresh on the shipped bytes (RTX 3090, sm_86).

• veizik run <model key> "prompt" --out DIR renders a video and writes its frames as PNG files to DIR/frames/; nothing is written inside the package.
• Serve: veizik serve <model> runs a resident render server that renders request after request in one process and one GPU context, with results bit-identical to veizik run.
• Four modes: quality (the reference path and the default), balanced, fast and ultrafast.
• An interrupted render continues from its last step with --resume, and the result is identical to a render that never stopped.
• Every file of the package is checked against its sealed manifest before each run.

cuda-qwen-1.1.4 · veizik Qwen (CUDA) v1.1.4 — Qwen3.5 / Qwen3.6, EOS-stop decode fix Current

Qwen3.6-35B-A3B and Qwen3.5 (0.8B, 2B, 4B, 9B) on your own NVIDIA GPU — text generation, serving, and image and video understanding — through veizik's own native CUDA engine: no Python runtime, no CUDA toolkit, no cloud at inference time. Activate once per device; all included models run. You supply the weights; nothing leaves the machine.

• Image and video input in the package's Qwen3.6-35B-A3B engine: it reads one image (PNG or JPEG) or one video (2 to 256 frames as PNG or JPEG files, in file-name order) together with a text prompt. Pass --image FILE or --video DIR to veizik run qwen36_35b_a3b.
• veizik serve <model>: a resident, batched, OpenAI-compatible HTTP server on 127.0.0.1 with server-sent-event streaming (/v1/chat/completions, /v1/completions, /v1/models, plus health and metrics endpoints).
• API-key authentication for the server: --require-auth with --keyfile or --api-key; a request without the key answers HTTP 401.
• veizik run for the five models with five modes (quality, balanced, fast, ultrafast, low-memory); the prompt can come from text or a file.
• Native GPU code for sm_80 through sm_120a, with the integrity of the bundle checked at every start.

cuda-wan-1.2.91 · veizik WAN (CUDA) v1.2.91 — Wan 2.1 & Wan 2.2 video Withdrawn

Text-to-video and image-to-video with Wan2.1-T2V-1.3B, Wan2.1-T2V-14B, Wan2.2-T2V-A14B and Wan2.2-I2V-A14B on your own NVIDIA GPU. Veizik runs the models through its own native engine: no Python, no CUDA toolkit. You supply the weights; the frames are written as PNG files to the folder you name. It also runs a resident render server with a host-loop (bin/veizik-resident) that keeps one process and one GPU context across requests, and mid-render preemption. A maintenance cut over 1.2.8: the base text- and image-to-video renders are byte-identical (the engine's non-LoRA code is unchanged), distributed now as a single signed .vzpack.

• veizik run <model key> "prompt" --out DIR renders a video and writes its frames as PNG files to DIR/frames/; nothing is written inside the package.
• Serve: veizik serve <model> runs a resident render server that renders request after request in one process and one GPU context, with results bit-identical to veizik run.
• Four modes: quality (the reference path and the default), balanced, fast and ultrafast.
• An interrupted render continues from its last step with --resume, and the result is identical to a render that never stopped.
• Every file of the package is checked against its sealed manifest before each run.

cuda-wan-1.2.8 · veizik WAN (CUDA) v1.2.8 — Wan 2.1 & Wan 2.2 video Superseded

Text-to-video and image-to-video with Wan2.1-T2V-1.3B, Wan2.1-T2V-14B, Wan2.2-T2V-A14B and Wan2.2-I2V-A14B on your own NVIDIA GPU. Veizik runs the models through its own native engine: no Python, no CUDA toolkit. You supply the weights; the frames are written as PNG files to the folder you name.. It also runs a resident render server with a host-loop (bin/veizik-resident) that keeps one process and one GPU context across requests, and mid-render preemption — an urgent request interrupts a running render, is served, then the first render resumes.

• veizik run <model key> "prompt" --out DIR renders a video and writes its frames as PNG files to DIR/frames/; nothing is written inside the package.
• Serve: veizik serve <model> runs a resident render server that renders request after request in one process and one GPU context, with results bit-identical to veizik run.
• Four modes: quality (the reference path and the default), balanced, fast and ultrafast.
• An interrupted render continues from its last step with --resume, and the result is identical to a render that never stopped.
• Every file of the package is checked against its sealed manifest before each run.

cuda-qwen-1.1.3 · veizik Qwen (CUDA) v1.1.3 — Qwen3.5 / Qwen3.6, single-file .vzpack Withdrawn

Qwen3.6-35B-A3B and Qwen3.5 (0.8B, 2B, 4B, 9B) on your own NVIDIA GPU — text generation, serving, and image and video understanding — through veizik's own native CUDA engine: no Python runtime, no CUDA toolkit, no cloud at inference time. Activate once per device; all included models run. You supply the weights; nothing leaves the machine.

• Image and video input in the package's Qwen3.6-35B-A3B engine: it reads one image (PNG or JPEG) or one video (2 to 256 frames as PNG or JPEG files, in file-name order) together with a text prompt. Pass --image FILE or --video DIR to veizik run qwen36_35b_a3b.
• veizik serve <model>: a resident, batched, OpenAI-compatible HTTP server on 127.0.0.1 with server-sent-event streaming (/v1/chat/completions, /v1/completions, /v1/models, plus health and metrics endpoints).
• API-key authentication for the server: --require-auth with --keyfile or --api-key; a request without the key answers HTTP 401.
• veizik run for the five models with five modes (quality, balanced, fast, ultrafast, low-memory); the prompt can come from text or a file.
• Native GPU code for sm_80 through sm_120a, with the integrity of the bundle checked at every start.

cuda-qwen-1.1.2 · veizik Qwen (CUDA) v1.1.2 — Qwen3.5 / Qwen3.6, multimodal Withdrawn

Qwen3.5 (0.8B, 2B, 4B, 9B) and Qwen3.6-35B-A3B on your own NVIDIA GPU, on-device, now with image and video input for Qwen3.6-35B-A3B. Veizik runs the models through its own native CUDA engine: no Python runtime, no CUDA toolkit, no cloud at inference time. You supply the weights; nothing leaves the machine.

• Image and video input in the package's Qwen3.6-35B-A3B engine: it reads one image (PNG or JPEG) or one video (2 to 256 frames as PNG or JPEG files, in file-name order) together with a text prompt. Pass --image FILE or --video DIR to veizik run qwen36_35b_a3b.
• veizik serve <model>: a resident, batched, OpenAI-compatible HTTP server on 127.0.0.1 with server-sent-event streaming (/v1/chat/completions, /v1/completions, /v1/models, plus health and metrics endpoints).
• API-key authentication for the server: --require-auth with --keyfile or --api-key; a request without the key answers HTTP 401.
• veizik run for the five models with five modes (quality, balanced, fast, ultrafast, low-memory); the prompt can come from text or a file.
• Native GPU code for sm_80 through sm_120a, with the integrity of the bundle checked at every start.

2026.10.02-r10 · Veizik for Apple Silicon Current

The engines are rebuilt with stronger protection against reverse-engineering of the on-device engine. Nothing you type or see changes: generation is token-for-token identical to r6.1–r6.5 (decode byte-identical, argmax unchanged). The veizik command, activation and the model catalogue are unchanged.

• Engine hardening: the engines behind veizik run and veizik serve are rebuilt with stronger protection against reverse-engineering. Render is token-for-token identical to r6.1–r6.5 (decode byte-identical, same argmax), so outputs do not change; the binaries are no longer byte-identical to the earlier builds.
• The benchmarks board's speed and memory figures were measured on the r6.1–r6.5 cores and carry to r10 because an independent security review verified that the engines render identically and a steady-state speed probe is within measurement noise; each carried row says so.

2026.10.02-r6.5 · Veizik for Apple Silicon Withdrawn

veizik devices disable, veizik devices enable and veizik deactivate now work on a Mac — the activation server used to refuse them with target_required, although the r6.3 and r6.4 notes said they worked. The engines are byte-for-byte the r6.1–r6.4 engines; only the veizik command changed.

• veizik devices disable, veizik devices enable and veizik deactivate work on a Mac. Until 2026-10-10 the activation server refused them from a Mac with target_required; the server was fixed, so they now work with r6.3 and r6.4 as well, and this build's veizik matches the fix.
• veizik devices lists every machine on your licence with its status (active, released or revoked) and the day it was last seen, and marks the one you are on.

2026.10.02-r6.4 · Veizik for Apple Silicon Withdrawn

veizik devices and veizik deactivate no longer ask you to set your licence id, and devices now lists every machine with its status and when it was last seen. The engines are byte-for-byte the r6.1, r6.2 and r6.3 engines; only the veizik command changed.

• veizik devices and veizik deactivate no longer need you to set your licence id. A machine that is already activated, including one activated with an earlier release, asks the activation server for it with its own device key — no browser step, no new approval — and keeps it; a browser or key activation made with this release stores it straight away. veizik config set license.id still works and still takes precedence.
• veizik devices lists every machine on your licence with its status (active, released or revoked) and the day it was last seen, and marks the one you are on. It used to show only the first machine, with no status.

2026.10.02-r6.3 · Veizik for Apple Silicon Superseded

Activation without a browser, for a Mac you reach over SSH: create an enrolment key and a CLI password on your account page and give them to veizik activate. Adds veizik deactivate, and veizik devices disable|enable now work. The engines are byte-for-byte the r6.1 and r6.2 engines; only the veizik command changed, and browser activation works as before.

• Activate without a browser (SSH, remote Macs, CI): create an activation key under “Browser-free activation” on your account page and a password under “CLI password”, then veizik activate --key KEY. A key is one-use by default and valid for 7 days; a reusable key works on as many machines as you choose, up to 25, for up to 90 days. Keys can be revoked on the account page at any time.
• The CLI password is asked for at a hidden prompt by default. For scripts: --password-stdin (recommended) or --password-file PATH. --password PW also works but is visible to ps and your shell history, and the command warns about it. A key created without a password is used with --no-password.
• veizik activate --email ADDR uses your account e-mail and CLI password instead of a key, if you turn that on under “CLI password”; it may e-mail you a 6-digit code.
• The machine still proves it holds its own key — the Secure Enclave on a Mac — exactly as in a browser activation; the key and password only stand in for the browser approval. When your account has an e-mail address on it, every machine activated this way is reported to that address, and the account page tells you if yours has none. A wrong key, password or code gets one answer that does not say which part was wrong; after 10 failed tries in a row a key is locked until you unlock it on the account page.

2026.10.02-r6.2 · Veizik for Apple Silicon Withdrawn

Sampling options on veizik run. The engines are byte-for-byte the r6.1 engines; only the veizik command changed, and sampling uses the sampler the engines already contained.

2026.10.02-r6.1 · Veizik for Apple Silicon Withdrawn

A fix release over r6: the CUDA option-forwarding overflow r6 listed as a known issue is fixed, and veizik keys now names the file and the reason when it cannot create its lock instead of reporting that another keys command is running.

2026.10.02-r6 · Veizik for Apple Silicon Withdrawn

The same four Qwen model families on Apple Silicon, in a new package. The entitlement id map now ships inside the image, so the runtime and the licence server read the same file rather than two copies of it.

2026.10.02-r5 · Veizik for Apple Silicon Withdrawn

Veizik for Apple Silicon. One signed package that reads supported safetensors checkpoints directly and runs eight language models on the machine you already have. 16.1 MB installed.

• veizik run --ignore-eos: generate exactly the requested number of tokens, so a fixed-length run produces exactly that length.
• --tokens N now prints exactly N tokens and says so on the summary line.
• Faster long-context decoding on more Macs: Qwen3-1.7B takes the long-context path from 8K tokens on M2 Pro, M1 Max and M1 Ultra, and Qwen2.5-1.5B gains a faster path on M2 Pro.
• veizik serve adds a TLS gateway, per-tenant quotas, a multi-node router, API keys with scopes and single sign-on, a usage ledger, request ids, tool calls and a service command.
• veizik serve now keeps each model loaded in a resident engine: one load per server, not per request, with model and feature licences checked when it loads and on every request.

2026.10.02-r3 · Veizik for Apple Silicon Withdrawn

Veizik for Apple Silicon. One signed package that reads supported safetensors checkpoints directly and runs eight language models on the machine you already have. 8.7 MB installed.

• Fixes an upgrade hazard: with an older disk image still mounted, an install could silently run the previous version's installer.
• Eight models run from the original checkpoint. There is no conversion step and no second copy on disk.
• 8.7 MB installed — 8,695,691 bytes in 85 files, summed from the published disk image itself, not from a build tree.
• Signed with a Developer ID and notarised by Apple; the image carries SHA256SUMS over its other files.

2026.10.02-r2 · Veizik for Apple Silicon Superseded

Veizik for Apple Silicon. One signed package that reads supported safetensors checkpoints directly and runs eight language models on the machine you already have. 8.7 MB installed.

• Strengthens how the engine protects its own code. Nothing changed in how your prompts, your outputs or your account are handled.
• Eight models run from the original checkpoint. There is no conversion step and no second copy on disk.
• 8.7 MB installed — 8,712,203 bytes in 85 files, summed from the published disk image itself, not from a build tree.
• Signed with a Developer ID and notarised by Apple; the image carries SHA256SUMS over its other files.

2026.10.02 · Veizik for Apple Silicon Superseded

Veizik for Apple Silicon. One signed package that reads supported safetensors checkpoints directly and runs eight language models on the machine you already have. 8.8 MB installed.

• Eight models run from the original checkpoint. There is no conversion step and no second copy on disk.
• Long prompts generate faster than in the previous release on every Mac we measured. The figures are not on this site: they were taken on development binaries, and this site only publishes numbers measured on the package you can download.
• veizik serve no longer needs a model named. Leave it out and it hosts every model installed on the machine, routing each request by its "model" field; /v1/models lists what it is hosting.
• The licence is leased into memory for a window the server sets — 48 hours as of 2026-10-02. The first run of a session waits for it; the runs that reuse it start in well under a second and finish with the internet switched off.
• 8.8 MB installed — 8,844,235 bytes in 85 files, summed from the published disk image itself, not from a build tree.

2026.10.01 · Veizik for Apple Silicon Superseded

Veizik for Apple Silicon. One signed package that reads supported safetensors checkpoints directly and runs eight language models on the machine you already have. Package figures were not recorded for this version.

• Eight models run from the original checkpoint. There is no conversion step and no second copy on disk.
• Every published figure for this build is on the benchmarks page.
• The licence is leased into memory for a window the server sets — 15 minutes at release, 48 hours as of 2026-10-02. The first run of a session waits for it; the runs that reuse it start in well under a second and finish with the internet switched off.
• Package figures were not recorded for this version.
• Signed with a Developer ID and notarised by Apple; the image carries SHA256SUMS over its other files.