The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestHigurashi Ai4B
Higurashi AiAny-to-Any

Haru-4B Brings Full-Duplex Speech to Six Languages

A compact 4-billion-parameter model handles real-time, two-way voice conversation across six languages in a single system.

Nov 12, 2026
NotableOther
Haru-4B

Higurashi AI has released Haru-4B, a compact speech-to-speech model built for full-duplex conversation. At roughly 4 billion parameters, it is designed to listen and speak at the same time rather than trading rigid turns, and it supports six languages out of the box.

Full-duplex is the detail worth dwelling on. Most voice assistants still operate in a strict turn-based loop: you talk, the system transcribes, generates, and then speaks back. A full-duplex approach lets the model handle overlapping speech, interruptions, and backchannel cues like "mm-hm" the way people actually converse — a meaningfully different interaction model from the familiar push-to-talk pattern.

Why it matters

Speech pipelines have traditionally stitched together separate components for recognition, language modeling, and synthesis. Haru-4B folds those capabilities into one system, spanning:

  • Speech-to-speech conversation as the primary mode
  • Speech recognition (ASR) and synthesis (TTS)
  • Coverage across six languages

At 4B parameters, the model sits in a size range that is realistic to run outside a data center, which lowers the barrier for developers building interactive voice interfaces. It is worth noting the model ships under a non-standard license, so teams should check the terms on the model page before building on it.

Details such as context length, supported sample rates, and the specific language list were not fully specified at release. For now, Haru-4B reads as an early but ambitious entry in the push toward natural, low-latency voice AI that runs at modest scale.

Sources

  • higurashi-ai/Haru-4B

    Hugging Face

    Visit
OlderBria releases RMBG-3.0 for background removalBriaai · Image Editing · last weekNewerMeituan's LongCat Adds Audio-Driven Avatar VideoMeituan · Image → Video · last week

Get the model

Hugging Face

Specs

Parameters4B
Size10.5 GB
PrecisionBF16
ArchitectureHaruForConditionalGeneration
LicenseOTHER
Downloads7.3K
Likes3.9K

Can you run it?

Runs on a laptop — about 5.1 GB at 4-bit.

  • BF16 (as published)

    16 GB GPU (RTX 4080 / 5070 Ti) · 24 GB Mac

    12.4 GB
  • 8-bit

    8 GB GPU (RTX 4060) · 16 GB laptop or Mac

    7.5 GB
  • 4-bit

    8 GB GPU (RTX 4060) · 16 GB laptop or Mac

    5.1 GB
Your machine
GGUF builds
  • KV cache approximated from model size (no config metadata) at 8,192 tokens.

Based on Hugging Face weights metadata. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


Modalities

Speech → TextAny-to-AnyText → Speech

The Weekly Weights

Every open release that mattered, one email a week.

0 comments

No comments yet. Be the first to weigh in.

More in Any-to-Any

All Any-to-Any →
Sage S2.1 Flash
Internlm/ReasoningMulti-GPU server

InternLM's Sage S2.1 Flash targets scientific reasoning

The 320B mixture-of-experts model is built to reason over scientific evidence and act inside interactive environments.

Nov 15, 2026
LongCat 2.5
Meituan/Text / LLMDatacenter

Meituan Releases LongCat 2.5, an MIT-Licensed MoE

The Chinese delivery giant expands its open-model push with a multimodal, agentic mixture-of-experts system.

Nov 4, 2026
LFM3-Omni-2B
LiquidAI/Any-to-AnyRuns anywhere

Liquid AI's LFM3-Omni-2B brings any-to-any AI to the edge

The 2-billion-parameter model combines vision, audio and speech-to-speech in a package small enough to run on device.

Oct 21, 2026