On 2026-11-06, India-based 169Pi released Alpie Seed 9B (Apache 2.0): SFT with DoRA plus multi-objective GRPO post-training on Qwen3.5-9B, keeping text/image/video input and a 262K native context, with FP16 plus Q4/Q3/Q2 GGUF builds. Vendor-reported FP16 scores with thinking on: MMLU-Pro 82.54%, GPQA Diamond 84.32%, LiveCodeBench v6 72.99%, SWE-Bench Verified 61.20% (Mini-SWE-agent, no retries), OCRBench 90.30%. The team also reports offline inference on iPhone 17 Pro, Galaxy Z Flip8, Mac M2, and Windows CPU-only.
Key Takeaways
- ✓Scope: 9B dense vision-language model post-trained from Qwen3.5-9B (not pretrained from scratch), Apache 2.0
- ✓Training: SFT DoRA (rank 64) then GRPO on 8xH100 for ~3472 GPU-hours / 3034 steps; release checkpoint is the FP16 Phase-4 merge
- ✓Coding (vendor-reported): LiveCodeBench v6 72.99% vs Qwen3.5-9B 65.6%; SWE-Bench Verified 61.20% with Mini-SWE-agent and no retries
- ✓On-device: Q4/Q3/Q2 GGUF; authors report ~9.59 tok/s on iPhone 17 Pro, ~14.1 on Mac M2, ~1.95 on Windows CPU-only
- ✓Limits: training data mix is undisclosed; headline scores are FP16 only; comparison numbers come from other model cards, not a shared harness
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and Pain Point
Cloud inference brings latency, cost, and data-egress constraints, while phone-sized models often lack reasoning and coding depth. 169Pi asks whether a 9B multimodal model can stay useful while actually running on laptops and phones.
Architecture and Mechanism
Alpie Seed keeps the Qwen3.5-9B stack and chat template: 32 layers, 4096 hidden size, hybrid Gated DeltaNet plus Gated Attention, plus the inherited 27-layer ViT. Post-training has two stages—SFT with DoRA (rank 64 / alpha 128), then GRPO with a separate DoRA+RSLoRA adapter against a frozen SFT reference on a 4-bit NF4 base (~3472 GPU-hours), merged back to FP16. Deployment adds Q4_K_M / Q3_K_M / Q2_K GGUF builds.
Benchmarks and Results
Vendor-reported FP16 scores with thinking on, max_tokens 65536, temperature 0: MMLU-Pro 82.54%, GPQA Diamond 84.32%, AIME 2026 83.33%, LiveCodeBench v6 72.99%, SWE-Bench Verified 61.20% (Mini-SWE-agent, no retries), OCRBench 90.30%, HallusionBench 79.88%. Comparison figures for Qwen3.5-9B / Gemma 4 12B and peers come from public model cards, not a shared harness. Quantized builds lack the same headline suite.
Developer Guide
HF: 169Pi/Alpie_seed; GGUF: 169Pi/Alpie_Seed_GGUF; Ollama: ollama run 169pi/alpie-seed (q2/q3/vision tags). Transformers >=5.2 via AutoModelForImageTextToText; vLLM can use --reasoning-parser qwen3. Validate FP16 leaderboard claims separately from on-device quant builds, and treat SWE-Bench as scaffold-dependent.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.