OpenBMB (MiniCPM)

Thesis

OpenBMB's evidence cluster resolves into two parallel thrusts. First, a research/engineering bet on on-device efficiency: small and MoE multimodal/language models, speculative decoding (DSpark), and a C++ omni inference engine (llama.cpp-omni) P7P10P12. Second, a productization push into AI agents via PilotDeck and StaffDeck, which show a near-daily release cadence across a multi-developer team P2P13E47. Around these sit expanding research lines in mixed-modality retrieval (Trident), test-time memory (T2Mem), and deep-research agents (AgentCPM-Report) P1P8P14. The connective tissue is shipping Apache-2.0/open weights optimized for local and edge deployment, backed by a developer-ecosystem GTM (the MiniCPM Spark Program) W1W2.

Signal desks

Hiring — No cited evidence in this pack. No job postings, role titles, team/functions, or location data appear; the pack consists of model cards, repo metadata, release notes, and third-party write-ups.

Forks — No direct fork events are cited. llama.cpp-omni is explicitly Fork: no but is "built on llama.cpp" P12. Upstream tools and models being adapted instead include Qwen3-VL-2B-Instruct (Trident is a finetune of it) P8, LLaVA-UHD v4's intra-ViT early compression (used by MiniCPM-V 4.7) P9P10, and the LoCoMo and LongMemEval benchmarks (T2Mem evaluation) P14.

Releases —

• MiniCPM-V 4.7 shipped in two Apache-2.0 variants: a 35B-A3B MoE ("first MoE model in the MiniCPM-V series," 13+ tokens/s on iPhone 17 Pro Max in ~3GB) P10E5 and a 1B on-device model claiming a new performance–efficiency Pareto frontier P9E6.
• Paired DSpark speculative-decoding draft checkpoints released for both the 1B and 35B-A3B targets (5 draft layers, trained on 2M sequences) P7P11E11E12.
• llama.cpp-omni (C++, MIT) launched as an omni multimodal inference engine for MiniCPM-o 4.5, with full-duplex streaming P12E23.
• Trident mixed-modality retrieval shipped as a repo plus the Trident-Qwen3VL-2B embedding model (MIT) and Trident-Train dataset P8P17E13E22.
• PilotDeck (task-oriented agent platform, TypeScript, 4,868 stars) cut near-daily releases across September–October 2026, adding built-in Computer Use (Cua Driver MCP), desktop installers (RPM, Windows ARM64, Ubuntu DEB, macOS), permission hardening, and Slack/Telegram IM channels P3P4P5P13E47.
• StaffDeck tagged v1.0.0-preview P2E17.
• Broader model family in evidence: MiniCPM5-2B E1, MiniCPM5-1B E4, VoxCPM2 E2, MiniCPM-V-4.6 E3, and robotics checkpoints MiniCPM-RobotManip / MiniCPM-RobotTrack E7E8.

Talking —

• The org's own X post frames MiniCPM-V 4.7 as "35B MoE multimodal AI. On your phone. With just 3GB of memory," spotlighting Lookahead Expert Scheduling (LES) and Self-Replay Fine-Tuning (SRFT) for customization W6.
• Third-party coverage amplifies the same narrative: AI/TLDR highlights the first MoE vision model decoding at 13+ tokens/s on-device W4, and evalaitools frames MiniCPM5-2B as "a tiny local AI model built for real agents" W5.
• The MiniCPM Spark Program S1 concluded with 16 winning projects (including an ESP32-S3 flight radar using MiniCPM5-2B and VoxCPM2), signaling developer-ecosystem traction W1.
• Local Model Watch catalogues OpenBMB's 189 published models and small-model focus W2; Edgepedia notes OpenBMB is jointly initiated by THUNLP and ModelBest W3.

Shipping

What actually shipped, through inspectable public artifacts, falls into three buckets. Models: MiniCPM-V 4.7 (1B and 35B-A3B) with matching DSpark draft checkpoints P9P10P7P11; the Trident-Qwen3VL-2B embedding model with Trident-Train data P8E13E22; and a dense set of prior MiniCPM5 / VoxCPM2 / MiniCPM-V-4.6 / robotics weights E1E2E3E4E7E8. Inference infrastructure: llama.cpp-omni, a C++ engine splitting MiniCPM-o into GGUF modules (VPM/APM encoders plus LLM) for consumer-grade hardware P12E23. Applications: PilotDeck's rapid release stream (Computer Use, desktop packaging across Linux/macOS/Windows, IM channels) and the StaffDeck v1.0.0-preview tag P3P4P5P13P2E17. DSpark checkpoints ship with vLLM --speculative-config usage P7P11.

Research themes

• On-device MoE inference: MiniCPM-V 4.7 35B-A3B pairs Lookahead Expert Scheduling (LES), layer-wise quantization, and DSpark speculative sampling to run an open MoE model on-device P10; the DSpark checkpoints document draft layers, hidden-state layer targets, and acceptance-length evals P7P11.
• Efficient vision encoding: intra-ViT early compression from LLaVA-UHD v4 plus mixed 4x/16x visual token compression P9P10.
• Mixed-modality retrieval: Trident mitigates "modality preference" where text outranks images, via Multi-Positive View InfoNCE over text/image/fused views P8P17.
• Test-time memory: T2Mem defers memory construction to inference, extracting episode metadata offline and building query-specific memory on LoCoMo and LongMemEval P14.
• Deep-research agents: AgentCPM-Report is an 8B deep-research agent using a Writing-As-Reasoning Policy (WARP) that interleaves drafting and deepening P1.
• Customization/RL: Self-Replay Fine-Tuning (SRFT) is promoted for task customization while preserving general performance W6; a JustRL-II-base-model and an UltraData code classifier also appear in the release stream E39E41.

Hiring & scaling

No hiring evidence is cited in this pack — no roles, teams, locations, or JD language. The gap is proven by the composition of the evidence itself: every artifact is a model card, release, repo, or web summary rather than a job or org-chart signal P1P2W2. The closest scaling indicator is indirect: PilotDeck release notes attribute work to at least four distinct contributors (Kaguya-19, mssssss123, blueberrycongee, xhd0728) across desktop, permissions, UI, and packaging changes, implying an active multi-developer product team P3P4P5P19. Treat any hiring/location inference as unsupported beyond this.

Category implications

• Strategy: OpenBMB is differentiating on capability density and on-device deployment — repeatedly asserting Pareto-frontier and "on your phone" claims against Qwen-class baselines P9P10W4W6. This is a resource-efficiency wedge, not a frontier-scale compute race.
• Infrastructure: The portfolio points to real inference-systems investment: DSpark speculative decoding P7P11, llama.cpp-omni for C++/GGUF omni inference P12, and declared support for vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX P9W5. These create deployment and chip-partner dependencies (iOS/Android/HarmonyOS; Rockchip, Houmo, Intel, AMD, Arm) P9.
• Product: PilotDeck and StaffDeck show a commercialization/agent-product buildout — Computer Use via MCP, desktop installers, token-usage stats, and messaging channels P13P3P4P5E47P2. The cadence implies an application-layer GTM separate from the open-weights research line.
• Research: Retrieval, memory, and agentic reasoning are being formalized into papers and repos (Trident, T2Mem, AgentCPM-Report), suggesting RAG/agent/memory as durable research axes rather than one-off releases P8P14P1.
• Hiring: Cannot be inferred — no hiring data exists in this pack.
• GTM: Apache-2.0/MIT open weights plus a named developer Spark Program (16 winning projects) and curated "model list" coverage indicate a community/ecosystem GTM aimed at builders and edge/device developers W1W2W3. No revenue or vendor claims are supported by the cited evidence.

Traction highlights

• MiniCPM5-2B: ~1.25M downloads, 1,817 likes E1.
• MiniCPM5-2B-DSpark: ~1.08M downloads E9.
• VoxCPM2: ~960K downloads, 1,722 likes E2.
• MiniCPM5-1B: ~332K downloads, 1,174 likes E4.
• MiniCPM-V-4.7-35B-A3B: ~187K downloads, 1,087 likes E5.
• MiniCPM-V-4.6: ~174K downloads, 1,245 likes E3.
• PilotDeck repo: 4,868 stars E47; Trident repo 11 stars, T2Mem 4 stars, llama.cpp-omni 35 stars at capture E29E36E23.
• Ecosystem: 16 award-winning Spark Program projects across multiple countries W1.