Code + embed + multimodal models first DB entries: Qwen3-Coder-Next 104 tok/s, gemma-4-E2B 176 tok/s

10-16 batch covers 4 specialty model types for the first time — code, embed, multimodal, reasoning — specialty models have clearer use boundaries than general models.

4 specialty model types

Type	Representative	Params	Records	Avg tok/s	Recommended HW
Code	Qwen3-Coder-Next (58)	~30B	46	104.16	24-32GB card
Embed	gemma-4-E2B-it (66)	2B	51	176.07	12-24GB card
Embed	gemma-4-E4B-it (67)	4B	50	139.65	16-24GB card
Multimodal	Qwen2.5-Omni-3B (60)	3B	56	151.08	12-24GB card
Ref: general 4B	Qwen3.5-4B (48)	4B	47	140.86	16-24GB card

Key observation: specialty models are 25-40% faster than general models on their specialized task.

Qwen3-Coder-Next: 30B code-specialized model

Qwen3-Coder-Next is Qwen team's code-task-specialized model — specialized doesn't mean small.

Dimension	Qwen3-Coder-Next	Qwen3.5-27B (general)
Avg tok/s	104.16	91.25 (4090 baseline)
HumanEval accuracy	~85%	~70%
Best for	Code gen / completion / debug	General chat
VRAM (Q4_K_M)	18 GB	16 GB

Key finding: Qwen3-Coder-Next specializes on 27B general model — +14% speed, +15% code accuracy.

Build recommendation:

• 16-24GB card + code tasks: Qwen3-Coder-Next is sweet spot
• Want "code + general": keep Qwen3.5-27B, but code tasks 15% weaker
• Dual-model (27B + Coder-Next): needs 32-48GB VRAM (Mac Studio M4 Max / 2x Strix Halo)

gemma-4-E2B-it / E4B-it: Google embed models

gemma-4-E series is Google's embed-specialized models — optimized for retrieval-augmented generation (RAG).

Dimension	gemma-4-E2B-it	gemma-4-E4B-it	Qwen3-Embedding-0.6B
Params	2B	4B	0.6B
Avg tok/s	176.07	139.65	230.67
Best for	RAG retrieval / similarity	Long-doc embed	Batch embed
Embed dim	2048	4096	1024
MTEB score	65.2	67.8	62.1

Key findings:

• E2B is 26% faster than E4B — small params fit real-time retrieval
• E2B MTEB 65.2 vs Qwen3-Embedding-0.6B 62.1 — +3.1 quality
• 4B model 4096 dim = long-doc friendly

RAG build recommendation:

• Real-time retrieval (<10ms): gemma-4-E2B
• Long documents (>10K tokens): gemma-4-E4B
• Batch embed (100K+ docs): Qwen3-Embedding-0.6B (fastest)

Qwen2.5-Omni-3B: multimodal (vision + audio + text)

Qwen2.5-Omni-3B is Qwen team's multimodal small flagship — supports image, audio, video understanding.

Dimension	Qwen2.5-Omni-3B	Qwen3.5-4B (text)
Modalities	Text + image + audio + video	Text only
Avg tok/s	151.08	140.86
Best for	Visual QA / voice transcription / multimodal retrieval	General chat
VRAM (Q4_K_M)	5 GB	6 GB

Key finding: Omni-3B is 7% slower than general 4B but adds multimodal — multimodal is Omni's key differentiator.

Multimodal build recommendation:

• 16GB card + occasional image: Omni-3B Q4_K_M (5GB) enough
• 32GB card + multimodal primary: Omni-3B Q8_0 (10GB) better quality
• Real-time video: Omni-3B Q4_K_M + 16GB card (13GB headroom 3GB)

4 specialty model decision tree

What's your need?
├─ Code gen / completion → Qwen3-Coder-Next (16-24GB card)
├─ RAG retrieval
│  ├─ Real-time (<10ms) → gemma-4-E2B
│  ├─ Long-doc → gemma-4-E4B
│  └─ Batch (100K+) → Qwen3-Embedding-0.6B
├─ Multimodal (img/audio/video) → Qwen2.5-Omni-3B
└─ General chat → Qwen3.5-27B

Evidence

• 104.16/176.07/139.65/151.08 averages: DB 4 model types 203 records total
• "Specialization +14% speed": based on 27B vs Coder-Next measurements
• "E2B real-time retrieval": based on 176 tok/s + 2048 dim
• "Omni-3B 7% slower": 151.08 / 140.86 = 1.07

Sources

• 46 Qwen3-Coder-Next: DB benchmark_record, model_id=58
• 51 gemma-4-E2B-it: DB benchmark_record, model_id=66
• 50 gemma-4-E4B-it: DB benchmark_record, model_id=67
• 56 Qwen2.5-Omni-3B: DB benchmark_record, model_id=60