Pin official source and exact weights
Mutable model names are not immutable checkpoints.
Record exact 40-character SHA, file inventory, config and runtime.
Unpinned dependencies can change outputs.
HIGURASHI AI · FIELD VERIFIED
Primary sources last checked on 17 Nov 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Haru-4B as a 5.27B stored tensor values (Haru-4B publisher name)-parameter full-duplex audio dialogue open-weight model in the Full-duplex audio dialogue family, with Higurashi AI listed as the developer. The checked source repository provides 1 recognized weight artifacts, format evidence for Safetensors and a context window that is not declared in the checked standard sources. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares apache-2.0. Commercial use allowed under apache-2.0 conditions. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: not-disclosed-in-checked-repository; data disclosure status: not-disclosed-in-standard-metadata-or-obvious-model-card-section.
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-11-17
Higurashi AI's Haru-4B is an openly downloadable, pretrained full-duplex speech-to-speech foundation model designed to listen and speak in overlapping conversational turns. Its official repository exposes a source-revision-pinned Safetensors checkpoint. The 4B product name should not be confused with the 5.27B stored tensor parameter count from repository metadata. Open Model Weights has not run inference or independently measured latency. Apache-2.0 is declared in model metadata, but no separate license file was observed in this checked revision.
WHY IT MATTERS
Haru-4B targets natural simultaneous listening and speaking instead of a turn-by-turn speech-recognition, text-generation and speech-synthesis chain. The publisher describes interruptions, backchannels and overlapping audio as its design goals.
The technical release describes three synchronized streams for incoming speech, generated speech and accompanying text. They advance at 7.5 frames per second, giving a frame duration of about 133 milliseconds; architecture descriptions are publisher evidence.
The official higurashi-ai/Haru-4B repository exposes one recognized weight artifact named model.safetensors and accompanying tokenizer and configuration files. The pinned SHA identifies the exact inspected revision, not just a mutable model-family name.
MODEL-SELECTION VIEW
Haru-4B targets natural simultaneous listening and speaking instead of a turn-by-turn speech-recognition, text-generation and speech-synthesis chain. The publisher describes interruptions, backchannels and overlapping audio as its design goals.
This is a pretrained base model rather than a general instruction-following voice assistant. Production users must implement audio capture, interruption handling, policies and application-specific behavior; a public checkpoint is not a ready-made customer-service system.
The technical release describes three synchronized streams for incoming speech, generated speech and accompanying text. They advance at 7.5 frames per second, giving a frame duration of about 133 milliseconds; architecture descriptions are publisher evidence.
Haru-4B is named for its language-model foundation. The checked Safetensors metadata counts approximately 5.27 billion stored tensor parameters across the checkpoint. The source configuration did not establish an ordinary text-token context window, so no token limit is asserted.
The official higurashi-ai/Haru-4B repository exposes one recognized weight artifact named model.safetensors and accompanying tokenizer and configuration files. The pinned SHA identifies the exact inspected revision, not just a mutable model-family name.
Downloaded artifact size and reported model parameters are distinct measurements. Developers should confirm tokenizer, audio frontend, output head and inference code compatibility before benchmarking or concluding that particular hardware can run continuous audio.
Higurashi reports approximately 65 milliseconds of processing time on an Apple M1 Max machine under its tested pipeline conditions. Crucially, that figure excludes audio framing, which adds up to roughly one 133-millisecond frame; it is not a guaranteed total interactive response latency.
No independent Open Model Weights latency, accuracy, interruption quality or multilingual evaluation has been run. A fair deployment test must include microphone buffering, playback, hardware, concurrent streams, speaker accents and recovery from interruption.
The model repository declares Apache-2.0 in first-party metadata, which normally permits commercial use subject to license conditions. A separate standard license file was not verified in the pinned repository, so operators should preserve provenance and review associated software and dataset licenses separately.
Continuous speech can contain biometric and sensitive personal content. Local operation does not, by itself, establish compliance with privacy, consent, recording or retention rules; those duties belong to the deployed application.
Haru is noteworthy for researchers building interactive speech agents where overlap and interruption control matter more than offline text-transcription accuracy. A transcription-only or traditional text-chat benchmark would not assess its key property.
Pin the official checkpoint and its inference code, build a representative conversation set and measure wall-clock latency including framing, multilingual turn-taking, speaker overlap, dropout and safety behavior. Record hardware and audio devices for every result.
IMPLEMENTATION & EVALUATION
Source-grounded deployment checklist; no Open Model Weights inference or benchmark was performed.
Mutable model names are not immutable checkpoints.
Record exact 40-character SHA, file inventory, config and runtime.
Unpinned dependencies can change outputs.
Repository labels do not cover every downstream dependency.
Check original model license declarations, extra licenses and required notices.
Unlicensed redistribution, privacy violations or incompatible integrations.
Source size and publisher numbers are not peak runtime requirements.
Record hardware, software versions, memory, latency, inputs and failure modes.
OOM, unsupported operators and unstable behavior.
Publisher measurements are not independent model rankings.
Compare pinned baselines on held-out representative tasks, report negative results.
False attribution of vendor performance to OMW.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
No. Its publisher describes it as a pretrained full-duplex speech base; instruction tuning and application safeguards are separate.
Yes. The verified official repository contains model.safetensors at a pinned revision.
No. The inspected repository metadata counts roughly 5.27B stored tensor parameters; the product name does not replace that artifact inventory.
A standard text-token context value was not verified in the checked configuration. Streaming audio framing is a different unit.
No. It is a publisher-reported processing measurement that excludes audio framing and was not independently reproduced by OMW.
The first-party repository declares Apache-2.0, but no separate standard license file was observed in the checked revision.
03 · PRACTICAL TECHNICAL PROFILE
Inspect weight-memory estimates, model lineage and runtime evidence separately from promotional or untested performance claims.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial use allowed under apache-2.0 conditions
This section summarizes the checked repository evidence. It is not legal advice.
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
04 · EVIDENCE AND UPDATES
Inspect dated source checks and any differences between published claims and the evidence; these checks are not independent model benchmarks.
SOURCE-BASED VISUAL BRIEF · NOTABLE
Three evidence-linked visual summaries: parameter scale, illustrative precision memory and verifiable provenance. No performance ranking is implied.
Active parameter count not established.
Repository-derived or publisher-described parameter countFRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
05 · SOURCE RECORD
Full repository-level attributes, licensing, format details and provenance are available below without hiding or rewriting the original data.
01 · WEIGHTS
02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: apache-2.0 (allowed commercial-use classification; see original terms).
Context: Unknown or not declared.
BF16 weights: 10.54 estimated decimal GB, not runtime VRAM.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →