GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

OpenBMB's MiniCPM-o 4.5 Ships Full-Duplex Omnimodal at 9B — Beats Qwen3-Omni-30B-A3B on 12GB RAM

Every current AI voice interface works the same way: user speaks, model waits for the end of turn, model replies. OpenBMB’s MiniCPM-o 4.5 ships the first genuinely open alternative at production-ready scale.

The model runs on OpenBMB’s Omni-Flow framework, which treats interaction as a continuous stream on a single shared temporal axis. Video tokens, audio tokens, LLM hidden states, speech tokens, and waveform generation are all time-aligned — the model perceives input and generates output simultaneously rather than sequentially. There is no walkie-talkie round-trip.

Key Numbers

  • Parameters: 9B total
  • RAM requirement: Under 12GB — fits on a single consumer GPU
  • Architecture: Full-duplex omnimodal via Omni-Flow shared temporal axis
  • Benchmark: Outperforms Qwen3-Omni-30B-A3B on omni-modal capabilities and speech generation quality
  • Deployment: Edge-capable; MLX-compatible for Apple Silicon
  • Licence: Open, with weights, code, and technical report

The model is 3x smaller than the Qwen model it outperforms on speech quality. The efficiency gap comes from architecture, not compression: Omni-Flow avoids the serialisation overhead of traditional turn-based pipelines by eliminating the buffering step between perception and generation.

The Architecture Difference

Most “omnimodal” models process modalities independently and stitch outputs together at the end of a turn. MiniCPM-o 4.5 processes all modalities in locked time steps. The model can interrupt, back-channel, and update its own output mid-stream as new sensory input arrives — the same constraint structure that makes human conversation feel natural.

Thinking Machines Lab previewed a similar concept from a research angle earlier this month with TML-Interaction-Small, but that model remains proprietary. OpenBMB has shipped the equivalent as an open-weight release with full reproducibility.

Context

MiniCPM-o 4.5 lands as voice AI benchmarks are beginning to formalise. The Full-Duplex Bench v3 (FDB-v3) published earlier this year showed Gemini Live 3.1 going silent 22% of the time — a symptom of the turn-based ceiling. MiniCPM-o 4.5 is the first open model explicitly designed to break that ceiling.

For developers building on-device voice agents, the sub-12GB footprint changes the target hardware. An iPhone 17 Pro already demonstrated 40 tokens/sec with Gemma 4 E2B under MLX; MiniCPM-o 4.5 extends that calculus to real-time multimodal interaction.

The weights, code, and Omni-Flow framework technical report are available via OpenBMB’s GitHub.