The choice is stored only in a first-party preference cookie on this site.
Built: 2026-11-13T17:51:31Z
Login
Useful AI. Keep the keys.
Tools that work for you. On hardware you control.
Hangry Labs packages practical local-first software with its models and dependencies, then gives you both a browser interface and useful APIs. Less dependency archaeology. Fewer mystery cloud calls. Clearer trade-offs.
Local-firstYour prompts, voices and transcripts stay close.
Baked imagesModels and dependencies travel together.
Open interfacesBrowser UI for people, HTTP APIs for systems.
Honest limitsUseful hardware guidance without magic smoke.
How the performance labels workRTF = generation time ÷ spoken-audio time. RTF ≤ 1.00 keeps pace with playback. Lab figures name the measured GPU; upstream figures and unmeasured compatibility guidance stay visibly separate.
Recorded GPU · RTX 5070 Ti
Explore everything
Browse the current projects
Already know what you need? Open any project for its purpose, hardware notes, starter command, examples, and source.
Local song generation
YuE2
Generate stereo 48 kHz songs from lyrics and musical direction. Edit ABC scores, replay musical takes, keep a local song library, and export Opus, MP3, FLAC, WAV and other formats.
Choose it when: Lyric-driven songs, score editing, acoustic variations, and replayable local music experiments.
Compute: NVIDIA GPU and a CUDA 12.9.1-compatible driver. Download the two GGUF models separately (about 4.34 GB); mount them read-only and keep the output volume for generated results.
Memory fitTested on RTX 5070 Ti 16 GB. The 120-second variation suite sampled 12,591-14,161 MiB whole-device peaks, including desktop/driver use. The 12+ GB group is not a verified 12 GB minimum.
Speed / RTF2 Nov 2026, RTX 5070 Ti: three seeds per setting, eight steps and resident models. A single 120-second rendition took 24.62 s median pipeline time; four shared-composition variations took 41.33 s total. Cold loading, audio packaging and ASR are excluded; normal Studio model eviction is unchanged. This is not a musical-quality score.
Restore speech, separate two-speaker recordings, enhance both tracks, or convert audio without loading a model. Compare original and processed audio locally through the studio or job APIs.
Choose it when: Cleaning speech, separating two voices, comparing restoration results, and converting recordings locally.
Compute: Linux x86-64 Docker on CPU or NVIDIA GPU; GPU mode requires a CUDA 13-compatible driver. Full images include verified weights; tiny images download the selected CPU/GPU assets on first use.
Memory fit8 GB is a conservative selection tier, not a measured minimum. Input duration, loaded models and restoration/separation settings change memory needs.
Speed / RTF30 Oct CUDA 13 validation on Windows/RTX 5070 Ti: 7.704 s of speech restored in 77.15 ms median, and 23.112 s in 245.09 ms (three warmed calls each). These fixture timings exclude loading, transfer and browser latency; they are not a sustained-capacity benchmark.
A local 82M-parameter speech workspace with 173 voices across 11 languages, five model packs, streaming, multi-character dialogue, OpenAI-compatible APIs, and opt-in MCP tools.
Choose it when: Lightweight narration, multilingual applications, Chinese speech, multi-character dialogue and local agent speech.
Compute: CPU with optional NVIDIA acceleration. Model packs load on demand; rolling latest includes Magic Studio and SSML-H features beyond the stable v1.0 release.
Memory fit3 Oct RTX 5070 Ti run: 2,052 MiB process-reserved peak and 3,413.6 MiB whole-device peak. That 70-voice run does not qualify every current model pack or a smaller GPU.
Speed / RTF4 Oct 2026, RTX 5070 Ti: 15.60x real time (RTF 0.0641), 350 warmed requests across the then-current 70 voices and 11 languages. The expanded 173-voice catalog has not had a complete rerun.
A local speech workspace with eight model options across six languages, output editing, OpenAI-compatible and native APIs, SSML dialogue, and sentence-streamed MP3/PCM.
Choose it when: English, French, Chinese and Japanese speech, optional Spanish/Korean, deterministic model selection, and standard client integrations.
Compute: CPU or NVIDIA GPU. Full images bundle six model families across English, French, Chinese and Japanese; Spanish/Korean require optional downloads and review of their model terms. The smaller image bundles the three English families.
Memory fit8 GB is conservative selection guidance, not a measured minimum. No reviewed GPU-memory benchmark establishes a lower-capacity requirement.
Speed / RTFThe current repository publishes capabilities and samples, but no reviewed RTX 5070 Ti timing/RTF benchmark. No speed figure is inferred from other TTS products.
A local voice studio for 600+ languages with voice design, microphone/reference cloning, reusable profiles, SSML-H dialogue, streaming and OpenAI-compatible APIs.
Choose it when: Multilingual speech, expressive voice design, consented cloning, reusable characters and structured dialogue.
Compute: GPU recommended, CPU fallback available. The 28 Oct FP32 CPU matrix recommended 2-6 GB RAM for short 140-character requests; newer snapshots default to FP16. Precision, reference length and ASR change usage. Default model weights are noncommercial; additional component terms apply.
Memory fit28 Oct RTX 5070 Ti: warmed short-reference runs peaked at 2,120.1 MiB process allocation and 4,864 MiB whole-device usage. Longer references, ASR and cold loading need separate headroom; 6/8 GB tiers are guidance, not validated minima.
Speed / RTF28 Oct 2026 clean-restart speech run: RTF 0.3103 across 180 requests in 18 languages on RTX 5070 Ti, averaging 1.301 s per request. This does not qualify every advertised language.
A local Qwen3-TTS workspace for voice design, consented cloning, built-in speakers, multilingual streaming, SSML-H dialogue and reusable voices/scripts with OpenAI-compatible APIs.
Choose it when: Expressive voices, consented cloning, multi-character scripts, streaming applications and reusable speech workspaces.
Compute: Supported NVIDIA GPU on Windows or Linux. Full images include five pinned checkpoints; tiny images download selected models into the persistent volume. Concurrent model residency needs more memory than the isolated-checkpoint tests.
Memory fit25 Oct RTX 5070 Ti suites: 0.6B whole-device peaks 6,341-6,666 MiB; 1.7B peaks 8,032-9,310 MiB. These include a 2,393 MiB desktop baseline and one checkpoint at a time; 8/12 GB tiers are guidance.
Speed / RTF25 Oct 2026: 0.6B complete responses reached 3.29-3.30x real time; 1.7B reached 2.86-2.94x. Streaming median first decoded audio was 304-342 ms on RTX 5070 Ti, excluding browser playback buffering.
A local VoxCPM2 workspace for 30-language speech, voice design, reference/transcript-guided cloning, reusable profiles, progressive MP3 streaming and HTTP APIs.
Choose it when: Creative voices, realistic consented cloning, multilingual output and reusable reference voices.
Compute: Supported NVIDIA GPU on Windows or Linux with the Nano-vLLM runtime. Run one model-serving container per GPU; keep concurrency one when fixed-seed repeatability matters.
Memory fit29 Oct RTX 5070 Ti: 7,873 MiB peak above a 1,934 MiB idle baseline, or 9,807 MiB whole-device usage. The delta is not total VRAM. 12 GB is conservative guidance.
Speed / RTF29 Oct 2026, RTX 5070 Ti: 312 warmed calls over 24 common languages at 10 steps. Median RTF 0.262, p95 0.317, and aggregate speed 3.85x real time. ASR transcript comparisons are not a perceptual speech-quality grade.
Local 30-language transcription, bounded live-microphone sessions, word/segment timestamps, subtitles, OpenAI-compatible APIs and opt-in MCP transcription from files or URLs.
Choose it when: Transcription, captions, meeting recordings, voice interfaces and applications using OpenAI-shaped APIs.
Compute: NVIDIA GPU; rolling latest pins vLLM 0.32 and CUDA 13, while stable v1.0 used vLLM 0.26. Start with 0.6B; the 1.7B model and optional forced aligner need more memory.
Memory fitHistorical vLLM 0.26 startup readings: 4,620 MiB (0.6B), 7,060 MiB (1.7B). First forced-aligner load added about 3.3 GiB. These are not current vLLM 0.32 peak-memory guarantees.
Speed / RTF2 Oct 2026 vLLM 0.26/0.6B baseline: 300 requests in 30 languages, 1,320.936 s of audio in 32.376 s, or 40.80x real time on RTX 5070 Ti. The rolling vLLM 0.32 runtime has no equivalent published rerun.
A Go dashboard for llama.cpp slot activity, token throughput, queueing, KV-cache reuse and request timelines, with a typed backend API.
Choose it when: Understanding an existing llama.cpp server, diagnosing capacity and seeing how requests share limited compute.
Compute: Hardware-neutral companion, not an inference engine. It connects to an existing llama.cpp server started with --metrics. The published release is v0.5.0; development continues at v0.6.0-SNAPSHOT.
Memory fitNo separate inference-model allocation in the dashboard. Model and KV-cache requirements belong to the llama.cpp server it monitors.
Speed / RTFAudio RTF does not apply to this monitoring dashboard; no inference-speed claim is made.
Start with the outcome. Hardware and VRAM are optional refinements, so 'not sure' is a perfectly useful answer.
1
Choose a category
2
What will run it? optional
Memory tiers are estimates, not tested minimums. Model, workload and other GPU use affect requirements. Memory and speed figures describe different test workloads, not a cross-product ranking.
3
How much GPU memory? optional
Choose a category to begin
0 matching projects
Pick TTS, STT, Other, or Docker Help above. We will reveal only the next decision that matters.
Local song generation
YuE2
Generate stereo 48 kHz songs from lyrics and musical direction. Edit ABC scores, replay musical takes, keep a local song library, and export Opus, MP3, FLAC, WAV and other formats.
Choose it when: Lyric-driven songs, score editing, acoustic variations, and replayable local music experiments.
Compute: NVIDIA GPU and a CUDA 12.9.1-compatible driver. Download the two GGUF models separately (about 4.34 GB); mount them read-only and keep the output volume for generated results.
Memory fitTested on RTX 5070 Ti 16 GB. The 120-second variation suite sampled 12,591-14,161 MiB whole-device peaks, including desktop/driver use. The 12+ GB group is not a verified 12 GB minimum.
Speed / RTF2 Nov 2026, RTX 5070 Ti: three seeds per setting, eight steps and resident models. A single 120-second rendition took 24.62 s median pipeline time; four shared-composition variations took 41.33 s total. Cold loading, audio packaging and ASR are excluded; normal Studio model eviction is unchanged. This is not a musical-quality score.
Restore speech, separate two-speaker recordings, enhance both tracks, or convert audio without loading a model. Compare original and processed audio locally through the studio or job APIs.
Choose it when: Cleaning speech, separating two voices, comparing restoration results, and converting recordings locally.
Compute: Linux x86-64 Docker on CPU or NVIDIA GPU; GPU mode requires a CUDA 13-compatible driver. Full images include verified weights; tiny images download the selected CPU/GPU assets on first use.
Memory fit8 GB is a conservative selection tier, not a measured minimum. Input duration, loaded models and restoration/separation settings change memory needs.
Speed / RTF30 Oct CUDA 13 validation on Windows/RTX 5070 Ti: 7.704 s of speech restored in 77.15 ms median, and 23.112 s in 245.09 ms (three warmed calls each). These fixture timings exclude loading, transfer and browser latency; they are not a sustained-capacity benchmark.
A local 82M-parameter speech workspace with 173 voices across 11 languages, five model packs, streaming, multi-character dialogue, OpenAI-compatible APIs, and opt-in MCP tools.
Choose it when: Lightweight narration, multilingual applications, Chinese speech, multi-character dialogue and local agent speech.
Compute: CPU with optional NVIDIA acceleration. Model packs load on demand; rolling latest includes Magic Studio and SSML-H features beyond the stable v1.0 release.
Memory fit3 Oct RTX 5070 Ti run: 2,052 MiB process-reserved peak and 3,413.6 MiB whole-device peak. That 70-voice run does not qualify every current model pack or a smaller GPU.
Speed / RTF4 Oct 2026, RTX 5070 Ti: 15.60x real time (RTF 0.0641), 350 warmed requests across the then-current 70 voices and 11 languages. The expanded 173-voice catalog has not had a complete rerun.
A local speech workspace with eight model options across six languages, output editing, OpenAI-compatible and native APIs, SSML dialogue, and sentence-streamed MP3/PCM.
Choose it when: English, French, Chinese and Japanese speech, optional Spanish/Korean, deterministic model selection, and standard client integrations.
Compute: CPU or NVIDIA GPU. Full images bundle six model families across English, French, Chinese and Japanese; Spanish/Korean require optional downloads and review of their model terms. The smaller image bundles the three English families.
Memory fit8 GB is conservative selection guidance, not a measured minimum. No reviewed GPU-memory benchmark establishes a lower-capacity requirement.
Speed / RTFThe current repository publishes capabilities and samples, but no reviewed RTX 5070 Ti timing/RTF benchmark. No speed figure is inferred from other TTS products.
A local voice studio for 600+ languages with voice design, microphone/reference cloning, reusable profiles, SSML-H dialogue, streaming and OpenAI-compatible APIs.
Choose it when: Multilingual speech, expressive voice design, consented cloning, reusable characters and structured dialogue.
Compute: GPU recommended, CPU fallback available. The 28 Oct FP32 CPU matrix recommended 2-6 GB RAM for short 140-character requests; newer snapshots default to FP16. Precision, reference length and ASR change usage. Default model weights are noncommercial; additional component terms apply.
Memory fit28 Oct RTX 5070 Ti: warmed short-reference runs peaked at 2,120.1 MiB process allocation and 4,864 MiB whole-device usage. Longer references, ASR and cold loading need separate headroom; 6/8 GB tiers are guidance, not validated minima.
Speed / RTF28 Oct 2026 clean-restart speech run: RTF 0.3103 across 180 requests in 18 languages on RTX 5070 Ti, averaging 1.301 s per request. This does not qualify every advertised language.
A local Qwen3-TTS workspace for voice design, consented cloning, built-in speakers, multilingual streaming, SSML-H dialogue and reusable voices/scripts with OpenAI-compatible APIs.
Choose it when: Expressive voices, consented cloning, multi-character scripts, streaming applications and reusable speech workspaces.
Compute: Supported NVIDIA GPU on Windows or Linux. Full images include five pinned checkpoints; tiny images download selected models into the persistent volume. Concurrent model residency needs more memory than the isolated-checkpoint tests.
Memory fit25 Oct RTX 5070 Ti suites: 0.6B whole-device peaks 6,341-6,666 MiB; 1.7B peaks 8,032-9,310 MiB. These include a 2,393 MiB desktop baseline and one checkpoint at a time; 8/12 GB tiers are guidance.
Speed / RTF25 Oct 2026: 0.6B complete responses reached 3.29-3.30x real time; 1.7B reached 2.86-2.94x. Streaming median first decoded audio was 304-342 ms on RTX 5070 Ti, excluding browser playback buffering.
A local VoxCPM2 workspace for 30-language speech, voice design, reference/transcript-guided cloning, reusable profiles, progressive MP3 streaming and HTTP APIs.
Choose it when: Creative voices, realistic consented cloning, multilingual output and reusable reference voices.
Compute: Supported NVIDIA GPU on Windows or Linux with the Nano-vLLM runtime. Run one model-serving container per GPU; keep concurrency one when fixed-seed repeatability matters.
Memory fit29 Oct RTX 5070 Ti: 7,873 MiB peak above a 1,934 MiB idle baseline, or 9,807 MiB whole-device usage. The delta is not total VRAM. 12 GB is conservative guidance.
Speed / RTF29 Oct 2026, RTX 5070 Ti: 312 warmed calls over 24 common languages at 10 steps. Median RTF 0.262, p95 0.317, and aggregate speed 3.85x real time. ASR transcript comparisons are not a perceptual speech-quality grade.
Local 30-language transcription, bounded live-microphone sessions, word/segment timestamps, subtitles, OpenAI-compatible APIs and opt-in MCP transcription from files or URLs.
Choose it when: Transcription, captions, meeting recordings, voice interfaces and applications using OpenAI-shaped APIs.
Compute: NVIDIA GPU; rolling latest pins vLLM 0.32 and CUDA 13, while stable v1.0 used vLLM 0.26. Start with 0.6B; the 1.7B model and optional forced aligner need more memory.
Memory fitHistorical vLLM 0.26 startup readings: 4,620 MiB (0.6B), 7,060 MiB (1.7B). First forced-aligner load added about 3.3 GiB. These are not current vLLM 0.32 peak-memory guarantees.
Speed / RTF2 Oct 2026 vLLM 0.26/0.6B baseline: 300 requests in 30 languages, 1,320.936 s of audio in 32.376 s, or 40.80x real time on RTX 5070 Ti. The rolling vLLM 0.32 runtime has no equivalent published rerun.
A Go dashboard for llama.cpp slot activity, token throughput, queueing, KV-cache reuse and request timelines, with a typed backend API.
Choose it when: Understanding an existing llama.cpp server, diagnosing capacity and seeing how requests share limited compute.
Compute: Hardware-neutral companion, not an inference engine. It connects to an existing llama.cpp server started with --metrics. The published release is v0.5.0; development continues at v0.6.0-SNAPSHOT.
Memory fitNo separate inference-model allocation in the dashboard. Model and KV-cache requirements belong to the llama.cpp server it monitors.
Speed / RTFAudio RTF does not apply to this monitoring dashboard; no inference-speed claim is made.
You do not need to know which installer or GPU package you need. Answer two questions and we will give you the safest starting path from the official documentation.
123
1
Which operating system do you use?
2
Which graphics hardware do you have? “Not sure” is fine
03 / Your setup path
Safe first checkConfirm Docker can start a tiny test container.
docker run --rm hello-world
Docker says “Hello from Docker!”?Perfect. Your runtime works. Next, choose a lightweight project and follow its repository instructions.