Daily Model Scout Report — 2026-04-14

Scouting VLMs released or trending in the last ~7 days that could beat or complement our current best garment classifier (qwen3-vl-8b-sft+grpo @ 0.9131 weighted on the 3,500-sample hard eval set).

All benchmarks referenced below are from the model providers' reports, not our hard eval — any claim of "may beat ours" requires a run through our 9-field JSON pipeline to confirm.

Tier 1 — Benchmark immediately (High)

1. Google Gemma 4 family — released 2026-04-02 (12 days ago)

• Variants: google/gemma-4-E2B, E4B, 26B-A4B (MoE, 8/128 experts), 31B (dense)
• License: Apache 2.0
• Modality: Text + Image (+ audio on E-series), 256K context, variable image token budgets (70→1120), function calling
• Why it may beat ours:
  • Brand-new multimodal base we have not touched. Prior Gemma/Granite family members (Granite4-Vision-SFT) scored 0.88 on our 100-sample eval — above our best Qwen3.5-2B SFT. A fresh, stronger base from the same lineage plausibly lifts that further after SFT+GRPO.
  • Native function calling makes strict JSON schema adherence cheaper to train, which is our hardest failure mode (closure, neckline, brand).
  • MoE 26B-A4B activates ~4B params per token → fits fits comfortably on the RTX PRO 6000 98GB and may match 8B dense cost.
• Fit for our stack: E4B (~4.5B effective) and 26B-A4B are the two obvious candidates. E4B replaces our 2B tier; 26B-A4B competes with Qwen3-VL-8B on inference cost while giving more capacity.
• Link: https://huggingface.co/google/gemma-4-E4B · https://huggingface.co/google/gemma-4-26B-A4B-it

2. Tencent Penguin-VL-8B / 2B — released 2026-03 (trending this week)

• Sizes: 2B, 8B (+ Penguin-Encoder standalone)
• License: Apache 2.0
• Architecture: Novel — vision encoder is initialized from a text LLM (Qwen3-0.6B) with bidirectional attention + 2D-RoPE, replacing CLIP/SigLIP. Language backbone is Qwen3-8B.
• Why it may beat ours: Author-reported head-to-head vs Qwen3-VL-8B shows Penguin-VL ahead on most image + reasoning benchmarks:
  • InfoVQA 86.8 vs 83.1, ChartQA 90.5 vs 89.6, AI2D 86.1 vs 85.7, RealWorldQA 75.8 vs 71.5, MathVista 77.4 vs 77.2, NextQA 85.4 vs 82.3
  • Qwen3-VL wins on OCRBench (896 vs 852) and DocVQA (tied). Our hard samples are garment-attribute reasoning, not OCR-heavy, so the Penguin trade actually favors us.
• Fit for our stack: Drop-in 8B replacement for the SFT+GRPO pipeline; same backbone lineage (Qwen3-8B) means our existing LoRA recipe should port with minimal tuning.
• Link: https://huggingface.co/tencent/Penguin-VL-8B · https://huggingface.co/tencent/Penguin-VL-2B

Tier 2 — Worth watching (Medium)

3. Qwen3-VL FP8 variants — updated 2025-11-26

• Full-collection FP8 pass dropped for 4B / 8B / 30B-A3B / 32B / 235B Instruct and Thinking variants.
• Why watch: Our NVFP4 quant of qwen3-vl-8b-sft+grpo loses ~1.9 points (0.9131 → 0.8945). Official FP8 weights from Qwen may close that gap further with less calibration work than NVFP4 required.
• Action: Quick A/B of Qwen3-VL-8B-Instruct-FP8 vs our NVFP4 on the same hard eval slice.
• Link: https://huggingface.co/collections/Qwen/qwen3-vl

4. LiquidAI LFM2.5-VL-450M — updated 2025-11-28, trending now

• Architecture: LFM2.5-350M LM + SigLIP2 NaFlex vision encoder; native 512×512 tiling
• License: lfm1.0 (check commercial terms before shipping)
• Why watch: Claims to beat SmolVLM2-500M across MMBench/MMStar/MMMU. Only viable if we ever need an edge/browser-tier model below our 2B floor — today our 2B models at 0.89 would almost certainly outperform, so this is for roadmap, not replacement.
• Link: https://huggingface.co/LiquidAI/LFM2.5-VL-450M

5. Qwen3-VL-Embedding-8B — recent addition to Qwen3-VL collection

• Why watch: Not a generator, but could power garment retrieval / nearest-neighbor de-duplication / hard-sample mining for the next training round.
• Link: https://huggingface.co/Qwen/Qwen3-VL-Embedding-8B

Tier 3 — Tangentially relevant (Low)

Recommended next actions

1. Spin up a Gemma 4 E4B SFT run on our existing ORR dataset using the proven SFT→GRPO→GTPO recipe. Target: match or beat qwen3-vl-2b-sft-grpo-v9 (0.8948) at roughly the same active parameter budget.
2. Baseline Penguin-VL-8B zero-shot on the 3,500-sample hard eval before committing any training time. If zero-shot is already ≥ Qwen3-VL-8B-Instruct base (0.78 territory on the 100-sample eval), queue a full SFT+GRPO cycle.
3. Swap in official Qwen3-VL-8B-Instruct-FP8 as the quant baseline and compare against our NVFP4 artifact.
4. Defer: LFM2.5-VL-450M (sub-2B is not a current deployment target), EXAONE-4.5 (license), HY-Embodied (wrong domain), OCR specialists.

Generated by the hf-model-scout skill. Eval anchor: qwen3-vl-8b-sft+grpo _overall.weighted_score = 0.9131 on our 3,500-sample hard eval.