cd /news/large-language-models/model-mnn-pocket-0-3b-instruct · home › topics › large-language-models › article
[ARTICLE · art-170975] src=quantized.uk ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Model: MNN-Pocket 0.3B Instruct

Alibaba MNN released MNN-Pocket 0.3B Instruct, a compact Chinese/English instruction model whose official BF16-derived Q8_0 GGUF export stores 361,551,232 parameters, according to the model's Hugging Face listing under taobao-mnn. The publisher documents a 32K context window, though the config and GGUF headers allow 40,960, and the Q8_0 file is 0.39 GB (386,751,520 bytes). At Q8_0 and 4K context the model needs about 0.5 GB of VRAM, rising to roughly 0.9 GB at the full 32K, with no measured speed or perplexity loss recorded for the build.

read5 min views5 publishedNov 17, 2026
Model: MNN-Pocket 0.3B Instruct
Image: Quantized (auto-discovered)

0.36B Alibaba MNN Pocket

A compact Chinese/English instruction model from Alibaba MNN. This entry uses the official BF16-derived Q8_0 export with 361,551,232 stored parameters, including its separate output table. The QAT Q4_1 checkpoint has tied embeddings and is indexed separately. The publisher advertises 32K context; config and GGUF headers allow 40,960, but this entry keeps the documented 32K scope. No measured speed or perplexity loss is recorded.

33K

Max Context

1

Quant Variants

GGUF Q8_0

Best Quality

—

Perplexity rise (best build)

Quantization Variants #

Per-quant VRAM, quality loss, and inference speed on RTX 4090

Measured = run on this site · Estimated = not run here: calculated, or a figure this site has not verified · Community = public reports

Which file each size is for

A file size is not a memory figure: the VRAM column is the calculator's total at 4K context — weights plus KV cache plus a 10% activation buffer. Units differ on purpose: file sizes are decimal GB, as Hugging Face lists them; the VRAM column is GiB (1024³ bytes) with 2% added to the weights, so 7.12 GB of file counts as about 6.8 GiB there. A vision projector (mmproj), where the repo ships one, is a separate download and is in neither number.

Related community runs

These runs describe their own settings. Their speeds do not replace this estimate or prove your configuration fits.

Our reviewed link index has no run for this model yet. You can browse community reports or share your own test below.

Browse more runs on llamaPerf ↗

Similar models #

[Compare with MNN-Pocket 0.3B](https://quantized.uk/tools/compare/?a=mnn-pocket-0-3b&b=mnn-pocket-0-3b-qat)

[0.30BConsumer GPUMac / Apple SiliconQ4_10.3 GB—perplexity rise](https://quantized.uk/quant-hub/mnn-pocket-0-3b-qat/)

MNN-Pocket 0.3B QAT

Alibaba MNN Pocket

The official QAT checkpoint converted to GGUF Q4_1 while preserving its trained 4-bit codes. Its tied input/output embeddings leave 302,830,976 stored parameters; this is a distinct checkpoint from the BF16-derived Q8_0 entry. All 28 layers use full GQA attention. Context is conservatively limited to the publisher-documented 32K despite 40,960 in the config and GGUF header. Task benchmark differences are not converted into a perplexity-loss percentage.

3BConsumer GPUMac / Apple SiliconQ4_K_M2.5 GB+3.5%perplexity rise

Llama 3.2 3B Instruct

Meta Llama 3.2

Tiny but capable. Runs on 4GB VRAM or 8GB RAM, even on phones via llama.cpp.

3BConsumer GPUMac / Apple SiliconQ4_K_M2.1 GB+3.2%perplexity rise

Qwen2.5 3B Instruct

Alibaba Qwen2.5

Tiny Qwen2.5 for edge devices. Runs on 4GB VRAM or Raspberry Pi class hardware.

1BConsumer GPUMac / Apple SiliconQ4_K_M0.9 GB+4.2%perplexity rise

Llama 3.2 1B Instruct

Meta Llama 3.2

Ultra-light Llama for mobile and embedded. Sub-2GB VRAM with Q4.

Running MNN-Pocket 0.3B Instruct locally #

At Q8_0 and 4K of context, MNN-Pocket 0.3B Instruct needs about 0.5 GB — 0.4 GB of weights, 0.05 GB of KV cache and a 0.0 GB activation buffer. The smallest cards in this index that clear that comfortably are 8 GB ones — 11 of them, from the Mac M1 8G up — and 78 of the 78 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs.

What longer context costs #

Going from 4K to 32K adds about 0.39 GB, taking the total to 0.9 GB. Weights do not move with context — only the KV cache does, and it grows linearly, so this is the number to watch when planning for long documents. The model's native window is 32K; holding all of it at Q8_0 would need about 1 GB.

Which build to download #

This index only tracks MNN-Pocket 0.3B Instruct in GGUF, across 1 level (Q8_0). No per-level perplexity sweep has been published for this model, so the quality column is empty rather than estimated — within one model, more bits per weight is the only ordering the data supports. No throughput has been measured for this model on this site's hardware. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 — CUDA only, and now archived, so local use rather than serving.

Common questions #

  • How much VRAM does MNN-Pocket 0.3B Instruct need?
  • About 0.5 GB at Q8_0 with 4K of context and batch 1, rising to roughly 0.9 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card.
  • Will MNN-Pocket 0.3B Instruct run on a 8GB GPU?
  • Yes — at Q8_0 and 4K context it needs about 0.5 GB, which leaves 5.5 GB spare on a Mac M1 8G, or on any of the other 8 GB cards here (11 in all). That is the smallest size in this index that clears it comfortably; 78 of 78 do.
  • Which quantization of MNN-Pocket 0.3B Instruct should I use?
  • Q8_0 is the only level in the index for this model, so there is nothing to choose between here.

Where this model fits #

Sized at Q8_0 with a 4K context window, smallest card first. Comfortable means the estimate uses at most 88% of the memory.

How to actually run this #

Deployment guides for this model and this class of hardware.

This page's figures change when the model or the runtime does.Last updated 2026-11-17 RSS → /feed.xml

── more in #large-language-models 4 stories · sorted by recency
── more on @alibaba mnn 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/model-mnn-pocket-0-3…] indexed:0 read:5min 2026-11-17 · —