MiniCPM5-2B Ranks Global #1 Under 4B with Day-0 Support on 9 Chips: ModelBest Pushes On-Device Agent Foundation to the Lowest Cost

Source: ModelBest / Jiazi Guangnian URL: https://mp.weixin.qq.com/s/rjFxrUylyGMqa5QtgypCdw Date: 2026-07-19 (WAIC 2026)

The Announcement

On July 19, ModelBest (Mianbi) together with OpenBMB released the on-device model MiniCPM5-2B, codenamed "Little Cannon," at WAIC 2026:

• AA-Index leaderboard: ranks global first among models under 4B with 54.26
• 9 chips Day-0 adaptation: Huawei Ascend, Hygon, Kunlunxin, MetaX, Moore Threads, T-Head, Tsingwei, Iluvatar CoreX, and NVIDIA
• Hybrid thinking + 512K context + tool calling, deep search, and code generation
• Upcoming open source: model weights and per-chip adaptation images will be released on ModelScope and Hugging Face

Head-to-Head Scores

| Model | Params | AA-Index Avg | |---|---|---| | MiniCPM5-2B | 2B | 54.26 | | LFM2.5-8B-A1B (8B sparse) | ~1B active | 49.15 | | Qwen3.5-2B | 2B | 41.66 | | Qwen3-1.7B | 1.7B | 40.89 | | Gemma-4-E2B-it | 2B | 37.62 |

Note that LFM2.5-8B-A1B is Liquid's 8B sparse architecture (~1B active) scoring 49.15 — ModelBest's 2B dense beats an 8B sparse model. This is not a slogan about parameter efficiency; it is a measured leaderboard number.

Per-Task Results: Agent Capability Is the Real Focus

| Task | MiniCPM5-2B | Note | |---|---|---| | AA-LCR long-context | 42.33 | 16 points ahead of runner-up | | τ²-Bench Telecom Agent | 90.35 | 4.97 points ahead of LFM2.5-8B-A1B | | AIME-2026 math | 60.95 | Qwen3-1.7B 36.88 / Gemma-4-E2B-it 32.71 | | ZebraLogic reasoning | 69.60 | First among compared models |

τ²-Bench Telecom evaluates customer-service agent scenarios; a score of 90.35 means a 2B on-device model can independently handle telecom-operator-level multi-step workflows. This matters directly for on-device agent deployment — AI assistants on phones and cars can perform real "procedural tasks," not just Q&A.

Day-0 Support on 9 Chips: Cross-Architecture Cloud-to-Edge Coverage

Through the unified software stack of the FlagOS community (vLLM-plugin-FL + SGLang-plugin-FL), ModelBest completed synchronized adaptation on 9 AI chips:

| Category | Chips | |---|---| | Domestic AI compute | Huawei Ascend, Hygon, Kunlunxin, MetaX, Moore Threads, T-Head, Tsingwei, Iluvatar CoreX | | International GPU | NVIDIA | | Edge platform | ARM |

ModelBest explicitly states: after FlagOS optimization, some domestic chips match or even exceed NVIDIA baselines in inference performance per unit of compute — and the AA-Index evaluation was run with multi-platform alignment.

Strategic Partnership with Intel 18A: 35B on 180 TOPS

ModelBest also signed an MOU with Intel on "on-device LLM + chip platform co-optimization." Benchmark figure: an Intel 18A Core Ultra platform with 180 TOPS supports local deployment of 35B-parameter models. Running 35B on consumer PCs is a milestone for on-device NPUs in 2026. Edge adaptation with AMD, MediaTek, and Qualcomm is also in progress.

UltraX Data Governance: Row-Level Edits with a 0.6B Model

As the internet-data dividend slows, ModelBest proposed "hierarchical data governance technology UltraX": using a 0.6B small model to perform row-level insert, delete, and edit operations on L1 web data, upgrading data governance from "traditional cleaning" to "executable editing programs."

Full Agent Training Pipeline

• 200B-scale Agent Midtraining
• Million-level high-quality trajectories for Agent SFT
• Scalable Agent RL alignment

The 200B figure refers to mid-training dedicated to agent capability, directly supporting the positioning that "the model can serve as an on-device agent foundation for local AI assistants on phones, PCs, and smart cockpits."

Why It Matters

1. The AA-Index #1 is not marketing spin — it is a third-party public leaderboard; 54.26 beating an 8B-sparse LFM2.5 is verifiable. 2. τ²-Bench Telecom 90.35 is a key number for on-device agents — a 2B model completing telecom-operator-level multi-step flows means phone assistants are no longer just Q&A. 3. Day-0 adaptation on 9 chips — amid export controls and domestic substitution, this multi-chip ecosystem reduces NVIDIA dependency at the supply-chain level. 4. 512K context + hybrid thinking in a 2B model with native tool calling, deep search, and code generation — foundation capabilities for on-device coding agents. 5. UltraX data governance industrializes "row-level insert/delete/edit" — a first public framing of governance as a program, not cleaning.

Risks and Open Questions

• Open-source date TBD — the announcement only says "in the near future."
• AA-Index is a composite leaderboard — the 54.26 average covers 7–8 benchmarks; single-domain extremes (math/code) may not lead. Domain-specific leaderboards need checking.
• τ²-Bench is a single scenario — the 90.35 on telecom customer service cannot be directly extrapolated to dining, booking, or office workflows; broader agent benchmarks are needed.
• Real accuracy on 9 chips — the post says "overall aligned with NVIDIA platforms" but does not disclose the boundary conditions of "matching or exceeding in some scenarios."

One-Sentence Takeaway

A 2B model beating an 8B-sparse model, 90.35 on τ²-Bench Telecom, and Day-0 adaptation on 9 chips — ModelBest has driven the cost of an on-device agent foundation to a new low while pushing NVIDIA dependency down a notch. Qwen3.5 and Gemma 4 series are now formally outpaced in the under-4B class.

Source links:

• Original article: https://mp.weixin.qq.com/s/rjFxrUylyGMqa5QtgypCdw
• Jiazi Guangnian: https://so.html5.qq.com/page/real/search_news?docid=70000021_4606a5cc9aa37852