AnswersPublished16 sources
What MNN-Pocket-0.3B Is—and What Its Mobile Speed Claims Show
MNN Pocket 0.3B is a roughly 300 million parameter model designed for mobile apps; its INT4 MNN build is 173 MB, and the team reports 350 tokens/s decoding on iPhone 18 Pro. The MNN README dates the release to November 4, 2026; a news report about it appeared November 6.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- MNN Pocket 0.3B is a roughly 300 million parameter model designed for mobile apps; its INT4 MNN build is 173 MB, and the team reports 350 tokens/s decoding on iPhone 18 Pro.
- The MNN README dates the release to November 4, 2026; a news report about it appeared November 6.
- The available sources do not establish per inference battery use or provide comparable prefill and peak memory measurements.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Alibaba’s Taobao MNN team’s MNN-Pocket-0.3B, reported open-sourced on November 6, 2026, and how does this roughly 300-million-parame. Article summary: MNN-Pocket-0.3B is Taobao’s approximately 300-million-parameter language-model family, co-designed with the MNN inference engine for on-device use; its main Hugging Face model is instruction-tuned, and its INT4 deploymen. Topic tags: general, general web, academic. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake num
MNN-Pocket-0.3B is a compact language model from Alibaba’s Taobao MNN team, designed for Chinese- and English-language instruction following in mobile apps. The team reports a 173 MB INT4 deployment build and 350 tokens per second of decoding on iPhone 18 Pro. Those figures make the model’s on-device ambitions concrete, but they do not by themselves establish its battery cost, prefill speed or performance across other phones.16
34
What MNN-Pocket-0.3B is
The model has roughly 300 million parameters and was designed around mobile-app tasks and resource budgets, with the model architecture, training methods and MNN inference engine developed together. Alibaba’s MNN README describes Chinese- and English-language instruction following, support for a 32K-token context, and an Apache 2.0 release.16
The README dates the release to November 4, 2026. A news article about the release was published November 6, so that later date is the report date rather than the release date given in the README.9
16
What the reported size and speed mean
The team’s INT4 W4A16 quantization-aware-trained MNN build is listed at 173 MB. That is a deployment-build size; the provided information does not itemize its files or clarify whether the figure includes every runtime component an app would need.16
19
The model card reports up to 466 tokens per second decoding on an M5 iPad Pro using Metal. A technical-report summary also gives that figure. These are reported decode rates, not independent comparisons across identical devices and settings.19
34
| Device and runtime | Reported decoding rate |
|---|---|
| iPhone 18 Pro, MNN | 350 tokens/s |
| M5 iPad Pro, Metal | Up to 466 tokens/s |
Tokens per second describes decoding—the rate at which a model generates output tokens. It is not a complete measure of an app’s responsiveness: the available sources do not provide corresponding prefill figures or enough test details to compare the two device results directly.
What is—and isn’t—established about mobile constraints
The design is explicitly aimed at fitting mobile-app resource budgets. The small INT4 build and reported decode rates are relevant to that goal, but they are not a full accounting of an app’s deployment costs.16
19
In particular, the available sources do not establish a per-inference battery figure, a standardized peak-memory measurement, or total app integration overhead. The reported package size should not be treated as the model’s full runtime memory use. Nor do the cited results show how performance changes under different workloads or device conditions.
Available model formats
The Hugging Face model card identifies taobao-mnn/MNN-Pocket-0.3B as the general-purpose instruction-tuned model and lists Base and SFT checkpoints in the family.19 A separate GGUF repository offers F16, Q8_0 and Q4_1 exports for llama.cpp-compatible runtimes, with listed file sizes of 726 MB, 387 MB and 192 MB respectively.
20
The Taobao MNN collection also lists an MNN-Pocket-0.3B-DFlash2 repository. The available listing does not establish its configuration or verify a claimed throughput multiplier, so it should not be read as evidence of a particular speed-up.21
The practical takeaway
MNN-Pocket-0.3B is notable for packaging a roughly 300-million-parameter model into a reported 173 MB INT4 MNN build and for the team’s high decoding-rate claims on iPhone 18 Pro and M5 iPad Pro. That supports its positioning as an on-device model for app developers; it does not, on the evidence available here, prove a specific battery advantage or a complete memory and integration budget.16
19
34
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What MNN-Pocket-0.3B Is—and What Its Mobile Speed Claims Show"?
MNN Pocket 0.3B is a roughly 300 million parameter model designed for mobile apps; its INT4 MNN build is 173 MB, and the team reports 350 tokens/s decoding on iPhone 18 Pro.
What are the key points to validate first?
MNN Pocket 0.3B is a roughly 300 million parameter model designed for mobile apps; its INT4 MNN build is 173 MB, and the team reports 350 tokens/s decoding on iPhone 18 Pro. The MNN README dates the release to November 4, 2026; a news report about it appeared November 6.
What should I do next in practice?
The available sources do not establish per inference battery use or provide comparable prefill and peak memory measurements.