Is MNN-Pocket-0.3B state of the art?Pending benchmarks.Announced Nov 4, 2026; no independent results yet.
The Alibaba Taobao MNN team's compact Chinese/English instruction-tuned text LLM, co-designed with its on-device MNN inference stack. Marketed as roughly 300M parameters (361.55M in HF safetensors metadata), with 32K deployed context and a lab-reported 173 MB INT4 W4A16 QAT MNN build. Apache-2.0 weights are public for local use; no public hosted API is confirmed.
The details so far
- Lab
- Alibaba Taobao MNN
- Announced
- Nov 4, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- ~300M advertised; 361.55M total safetensors parameters
- Team and lineage
- MNN Team, Taobao and Tmall Group of Alibaba; instruction-tuned from taobao-mnn/MNN-Pocket-0.3B-Base. This is a separate MNN team release, not attributed to the Qwen lab merely because a qwen3 tag appears
- Architecture
- 28 layers, hidden size 896, GQA with 14 query / 2 KV heads, tied input/output embeddings, 65,536 vocabulary; BF16 native checkpoint
- Context
- 32K supported in deployment; configuration/card lists 40,960 positions, not a claim of 40K deployed support
- License
- Apache-2.0 verified in model card and LICENSE.txt
- Formats
- Public Transformers safetensors, GGUF and MNN formats. MNN INT4 W4A16 quantization-aware build is advertised at 173 MB; GGUF Q4_1 uses the QAT checkpoint
- Limitations
- Team does not recommend reasoning-heavy advanced mathematics or long-horizon code generation; latency/throughput depends on hardware, backend and context
- Related
- Separate from Base and intermediate SFT checkpoints, DFlash2 auxiliary variants, and the non-autoregressive Dec decision model. GGUF/MNN packaging variants stay on this instruction-model page
- Date evidence
- Official alibaba/MNN README release note dated Nov 4, corroborated by a public HF discussion that day. Sep 23 repo creation and Sep 26 release-labelled commit do not independently prove historical public access
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: INT4 W4A16 QAT deployment footprint is 173 MB. At 512-token context, decode reaches 466 tokens/s on iPad Pro M5 Metal (231 on its CPU); these are team measurements, not independent results.
Lab claim: average across the team's 11 Chinese/English benchmarks is 50.00 for native precision and 49.59 for INT4.
Sources
REPOMNN team model card and weightshuggingface.co/taobao-mnn/MNN-Pocket-0.3BREPOOfficial MNN README Nov 4 release newsgithub.com/alibaba/MNN · Nov 4, 2026REPOPublic HF discussion corroborating Nov 4 accesshuggingface.co/taobao-mnn/MNN-Pocket-0.3B/discussions/1 · Nov 4, 2026REPOMNN-Pocket Apache-2.0 licensehuggingface.co/taobao-mnn/MNN-Pocket-0.3B/blob/main/LICENSE.txtLABMNN team technical reporthuggingface.co/taobao-mnn/MNN-Pocket-0.3B/blob/main/MNN-Pocket-0.3B-Technical-Report.pdfTracking continues until an independent board lists it.