OpenBMB Releases MiniCPM5-2B With Apache-2.0 Weights and Agent Training Data

OpenBMB released MiniCPM5-2B on September 7, 2026, making its compact language model available with Apache-2.0 weights, multiple quantized formats, intermediate checkpoints, deployment documentation, and substantial portions of its training data.

The release is notable for combining a small deployment footprint with measurable agent capabilities. Artificial Analysis currently gives MiniCPM5-2B an Intelligence Index score of 15 under version 4.2, the highest result for an open-weight model below four billion total parameters. That independent result places it four points ahead of Granite 4.2 3B and level with the much larger Qwen3.5 9B reasoning configuration.

The score of 23 circulated in OpenBMB’s announcement and the original X post is not the model’s current index score. It came from Artificial Analysis Intelligence Index v4.1.1. The evaluation provider subsequently changed its benchmark mix and weightings for v4.2, reducing MiniCPM5-2B’s reported score to 15. Results from the two versions are not directly comparable.

1. What OpenBMB Released

MiniCPM5-2B is a dense, text-only causal language model based on the standard LlamaForCausalLM architecture. Despite its product name, the checkpoint contains 2,516,756,480 parameters in total, of which 1,981,982,720 are non-embedding parameters. Artificial Analysis therefore categorizes it as a 2.6-billion-parameter model.

The architecture has 42 layers, grouped-query attention with 16 query heads and two key-value heads, and a native context length of 131,072 tokens. Its standard Llama-compatible implementation means common inference frameworks can load it without a model-specific architecture fork or custom kernel.

OpenBMB released more than a single instruction-tuned checkpoint. The available artifacts include the final BF16 model, a pre-training-only base checkpoint, a mid-training checkpoint, an SFT-only checkpoint from before reinforcement learning and distillation, a four-bit GPTQ build, an Apple Silicon-oriented four-bit MLX build, and GGUF files for llama.cpp, Ollama, and LM Studio.

The official GGUF repository provides a 5.04 GB F16 file, a 2.68 GB Q8_0 quantization, and a 1.56 GB Q4_K_M quantization. These sizes make fully local operation practical on many consumer systems. OpenBMB presents the model as suitable for edge and on-device use, although the release materials do not provide independently verified phone-specific speed, memory, battery, or thermal measurements. A model file fitting in phone storage is not, by itself, proof of acceptable mobile performance across devices or long contexts.

There is also a DSpark draft model for speculative decoding. When paired with MiniCPM5-2B through SGLang, it proposes blocks of tokens that the target model verifies, with the goal of accelerating generation without changing the target model’s accepted output.

At publication time, Hugging Face lists no hosted inference provider for the primary checkpoint. The immediate access route is therefore downloading and self-hosting the model rather than purchasing calls to an official API.

2. The Benchmark Result—and the Important Version Change

OpenBMB reports an average score of 53.9 across 34 evaluations covering code, mathematics, instruction following, knowledge, long-context understanding, tool use, and several kinds of agents. In the developer’s comparison table, the highest average among the listed larger reference models is 51.1 for Qwen3.5-4B.

That 53.9 figure should be read as a developer-reported composite, not as a universal model rating. It averages different benchmarks and includes many results reproduced internally by OpenBMB. Selected reported scores include 69.1 on LiveCodeBench v6, 86.5 on AIME 2026, 66.6 on BFCL v4, 46.4 on SWE-bench Verified, 88.7 on the GAIA Text-103 subset, and 20.8 on τ³-Bench Banking. OpenBMB labels which entries came from Artificial Analysis and which were reproduced through its own evaluation setup.

Artificial Analysis provides the strongest independent confirmation of the model’s position within its size class. Under Intelligence Index v4.2, MiniCPM5-2B scores 15, compared with 11 for Granite 4.2 3B and an estimated 14 for reasoning-enabled Qwen3.5-4B. Ling 3.0 Tiny scores one point higher at 16 but has approximately three times as many total parameters.

The independent results also define the model’s limits. MiniCPM5-2B scored 9% on Humanity’s Last Exam, 9% on Terminal-Bench v2.1, 0% on CritPt, 26% on SciCode, and 59% on Artificial Analysis’s long-context reasoning evaluation. Those figures do not support treating the model as a general substitute for larger frontier systems.

Its agent results are more competitive. MiniCPM5-2B reached an Elo score of 831 on GDPval-AA v2, where 1,000 represents the human baseline used by the evaluation. It scored 21% on τ³-Banking and an Elo of 438 on AA-Briefcase. Artificial Analysis defines its Agentic Index as a weighted average of those three evaluations.

The model used approximately 19,000 output tokens per Intelligence Index task, including about 11,000 reasoning tokens. That tied Granite 4.2 3B for the lowest output-token use in the comparison set and was roughly one-third of Ling 3.0 Tiny’s 56,000 tokens. Token count is especially relevant for local deployment because longer reasoning traces increase latency, memory pressure, and energy consumption.

One unusual result requires caution: MiniCPM5-2B’s relatively favorable AA-Omniscience score came largely from abstention. It attempted only 29% of questions, producing a 78% non-hallucination rate but just 8% accuracy. The result indicates conservative answering behavior on that test, not broad factual mastery.

3. Why the Agent Capabilities Matter

MiniCPM5-2B is explicitly trained for tool use, coding, search, and multi-step agent workflows rather than being only a compact chat model. It emits tool calls in an XML-style format. OpenBMB recommends SGLang for function calling because its minicpm5 parser converts those outputs into OpenAI-compatible tool_calls.

The released agent instruction-tuning dataset makes this focus unusually concrete. UltraData-SFT-Agent-2609 contains 483,661 trajectories: 311,006 general-agent examples, 82,760 tool-use examples, 69,895 code-agent examples, and 20,000 search-agent examples.

These are multi-turn trajectories rather than isolated prompt-and-answer pairs. They can include tool definitions, calls, environment responses, verification steps, errors, retries, and final results. The covered tasks range from function calling and database operations to software engineering, web retrieval, file processing, office workflows, and multi-turn memory.

The dataset is not a complete collection of executable agent environments. OpenBMB states that it does not include the original environments, tool implementations, tests, or sampling code, and that many virtual APIs represented in the trajectories do not exist in real deployments. Researchers can inspect and reuse the interaction records, but they cannot automatically replay every trajectory from the released files alone.

For developers, the practical change is that a small local model can now be evaluated as the decision-making component of an agent without sending prompts, source code, or tool results to a remote model provider. Whether it is reliable enough for a particular workflow still depends on application-level testing, permissions, validation, and recovery logic.

4. The Training Release Goes Beyond Model Weights

OpenBMB describes MiniCPM5-2B’s training process in three broad stages: base training, mid-training, and post-training. Post-training then proceeds through supervised fine-tuning, reinforcement learning, and on-policy distillation.

The released UltraData-RL-2609 dataset contains 85,995 verifiable-reward samples. Its four categories are mathematics with 32,412 samples, code with 23,665, long-context tasks with 18,046, and scientific or knowledge reasoning with 11,872.

The reward mechanism differs by category. Mathematics and knowledge tasks use extractable reference answers; long-context samples require a supported answer within the supplied context; and code submissions are executed against test cases. OpenBMB says labels were checked through methods including multi-model consensus, answer comparison, and test-case validation, while tasks already solved consistently by the initialization model were removed.

OpenBMB trained specialized reinforcement-learning teachers for domains including mathematics, code, writing, and agent tasks. It then used on-policy distillation to combine 16 expert models—including five agent specialists—into the released checkpoint. At each response position, the method uses reverse KL divergence between student and teacher token distributions as its advantage signal.

According to OpenBMB’s ablation results, reinforcement learning plus on-policy distillation improved reasoning and general evaluations by an average of 10.96 points and agent evaluations by 6.96 points over the SFT checkpoint. These are developer measurements, but the availability of pre-RL and final checkpoints gives outside researchers a way to test the claimed improvement.

The model repository and weights use Apache 2.0. The accompanying datasets require additional care: their documentation says upstream licenses continue to apply and includes restrictions against unauthorized unchanged mirroring or commercial repackaging. Organizations planning to redistribute the data should review each dataset’s terms rather than assuming that the model-weight license governs every training artifact.

5. What Changes for Local AI Development

MiniCPM5-2B follows the 1B MiniCPM5 checkpoint released in May 2026, but it raises the useful capability ceiling without moving out of the consumer-device class. The 1.56 GB four-bit file, standard architecture, long context, and documented support for llama.cpp, Ollama, LM Studio, MLX, Transformers, vLLM, and SGLang reduce the integration work normally associated with a new model family.

The release is particularly relevant where data locality or intermittent connectivity matters. A downloaded checkpoint can operate without sending subsequent prompts to OpenBMB, while the Apache-2.0 model license permits modification and commercial use subject to the license’s conditions.

The strongest verified case is not that a 2B-class model now matches large systems on every task. It does not. The evidence instead shows that agentic tool use, coding, and structured reasoning can be compressed into a model with roughly two billion non-embedding parameters while remaining competitive with several larger open-weight models. The downloadable intermediate checkpoints and training datasets also make that result more inspectable than a benchmark claim attached only to final weights.

Frequently Asked Questions

Is MiniCPM5-2B really a two-billion-parameter model?

It has 2.52 billion total parameters and 1.98 billion non-embedding parameters. “2B” refers to its model class and non-embedding scale.

What is its current Artificial Analysis score?

MiniCPM5-2B scores 15 on Intelligence Index v4.2. The widely quoted score of 23 came from v4.1.1 and cannot be directly compared with the updated index.

Can it run without an internet connection?

Yes. Once its weights and runtime are downloaded, GGUF, MLX, GPTQ, or full-precision versions can be run locally. Actual speed and memory requirements depend on the hardware, quantization, and context length.

Is the full training process reproducible from the release?

Only partially. OpenBMB released staged checkpoints, recipes, and substantial SFT and RL datasets, but the agent dataset does not include all environments, tool implementations, tests, or sampling infrastructure.

Does MiniCPM5-2B support function calling?

Yes. It generates XML-style tool calls, and OpenBMB recommends SGLang’s minicpm5 parser for converting them into OpenAI-compatible tool-call objects.

Sources

• Original X post from @itsPaulAi
• OpenBMB MiniCPM GitHub repository and release documentation
• MiniCPM5-2B model card and weights
• MiniCPM5-2B GGUF files and quantization sizes
• Artificial Analysis: MiniCPM5-2B results on Intelligence Index v4.2
• UltraData-SFT-Agent-2609 dataset documentation
• UltraData-RL-2609 dataset documentation