NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

vivo's BlueLM-2.5-3B Lets a 3B Model Decide How Hard It Thinks

vivo's AI Lab released BlueLM-2.5-3B, a 2.9-billion-parameter on-device multimodal model that can switch between a fast direct answer and a longer reasoning pass — and even cap how many tokens it spends thinking.

2026-11-12 · 987 words · NeuroAI
vivo's BlueLM-2.5-3B Lets a 3B Model Decide How Hard It Thinks

A phone maker's model usually stays inside the phone. vivo, best known for Android handsets, has been publishing its BlueLM (蓝心大模型) research openly, and the latest entry is a small multimodal model built around a deceptively simple idea: let the model choose, per question, how much it wants to think.

Most reasoning models either always think or never do. BlueLM-2.5-3B, described in an arXiv technical report dated July 2025, is pitched as the first 3-billion-scale multimodal model that supports both a thinking mode and a non-thinking mode, with explicit control over the thinking budget. For a device that runs on battery, that switch is not a gimmick — it is the difference between a snappy reply and a drained phone.

Key takeaways

  • BlueLM-2.5-3B is a unified dense multimodal large language model from vivo AI Lab with 2.9 billion parameters (a 2.5B language model plus a ~400M vision encoder), designed for edge devices.
  • It is described as the first 3B-scale multimodal model to support both thinking and non-thinking modes, and to let the user cap the thinking-token budget via a special token.
  • In thinking mode it reaches text-only performance comparable to Qwen3-4B, and trails the larger Kimi-VL-A3B-16B by about 5% on average across multimodal evaluations.
  • In non-thinking mode it outperforms Qwen2.5-VL-3B on most multimodal benchmarks, using substantially less total training data than either comparison model.
  • Context length was extended from 4K to 32K tokens via the YaRN method; the technical report is arXiv 2507.05934.

Thinking as a dial, not a switch

Reasoning models have gotten good by spending thousands of tokens working through a problem before answering. That works beautifully on a server and terribly on a phone, where every token is battery and latency. BlueLM-2.5-3B treats thinking as a tunable budget rather than an on/off flag: a special token lets the system or the app decide whether to reason at all, and if so, how many tokens the reasoning pass is allowed to consume.

The design solves a real product problem. A camera app asking "what is this plant" wants a one-line answer now; a student asking "prove this geometry claim" wants the full chain. Forcing both through the same heavy path wastes power on the first and underserves the second. Giving the device control over the budget lets a phone behave like a phone for trivial queries and like a tutor for hard ones.

The edge-efficiency backbone

Behind the mode switch is a compact, efficiency-first build. The vision side uses a SigLIP2 vision transformer (about 400M parameters, 27 layers) with a dynamic-resolution processor that tiles images up to 1512×1512, so high-resolution input stays feasible on small hardware. The language backbone is a 2.5B transformer, and a two-layer adapter with spatial downsampling keeps token counts and compute down.

vivo reports the model was trained through careful data curation, key-data resampling, and hybrid heterogeneous reinforcement learning on a high-performance infrastructure. The headline efficiency claim is that BlueLM-2.5-3B matches or beats larger comparison models while trained on far less total data than Qwen2.5-VL-3B and Qwen3-4B — a data-efficiency story that matters as much as raw scores for a company shipping to millions of devices.

Its predecessor, BlueLM-V-3B, set the tone: a 2.7B language model with a 400M vision encoder that hit 24.4 tokens per second on a MediaTek Dimensity 9300 with 4-bit weight quantization, and led the OpenCompass average among models at or below 4B, surpassing MiniCPM-V-2.6 and InternVL2-8B. BlueLM-2.5-3B is the reasoning-capable successor to that line.

Where it actually lands

The comparisons are vivo's own, drawn against well-known open models. On text-only benchmarks in thinking mode, BlueLM-2.5-3B is pitched as comparable to Qwen3-4B; across multimodal evaluations it trails the much larger Kimi-VL-A3B-16B by roughly 5% on average. In non-thinking mode it beats Qwen2.5-VL-3B on most multimodal sets. None of those are independent leaderboard results yet — they are the report's positioning against established baselines.

For a global reader, the takeaway is less "vivo beat X" than "the on-device reasoning category is maturing." A handset maker is no longer content to license a cloud model; it is publishing reproducible research on how to make a 3B model reason selectively and run locally. According to Li An, chief scientist at BrainNet (脑机网), controllable thinking is the right unit of competition for edge AI, because what users feel is not peak benchmark score but whether the device answers fast when it should and thinks hard when it must.

Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

Honest limitations

  • The comparison figures — comparable to Qwen3-4B, ~5% behind Kimi-VL-A3B-16B, ahead of Qwen2.5-VL-3B — are from vivo's technical report, not an independent benchmark; third-party reproduction has not yet published competing numbers.
  • "First 3B-scale MLLM with thinking and non-thinking modes" is vivo's framing; the field moves fast, and similar small-model reasoning controls may appear from other labs.
  • A 2.9B model, even with a thinking mode, sits below the largest multimodal systems on the hardest reasoning and longest contexts; the 32K context helps, but it is not a substitute for scale on the toughest tasks.
  • The 24.4 tok/s figure for the predecessor is measured on a specific MediaTek Dimensity 9300 with 4-bit quantization; real-world speed on other chips will vary.
  • The thinking-budget control is only useful if the app or system sets it well; a poorly chosen budget defeats the efficiency purpose.

What readers can do now

  • If you build on-device assistants, prototype BlueLM-2.5-3B and wire the thinking-budget token to a difficulty heuristic — fast path for trivial queries, reasoning path for hard ones — instead of a fixed mode.
  • If you train small multimodal models, study the report's data-resampling and hybrid-RL recipe as a path to matching larger models on less data.
  • If you ship to phones, benchmark it on your target chip at 4-bit and measure real battery and latency per mode before deciding whether a cloud fallback is still needed.

Related coverage

Cite this article

NeuroAI (脑机网), "vivo's BlueLM-2.5-3B Lets a 3B Model Decide How Hard It Thinks," 12 November 2026, https://neuroai.site/a/na-model-vivo-bluelm-edge

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories