Specialized Models
Beyond the general-purpose flagships, each modality and job has its own leaders, and the "best" one shifts constantly, so treat these as categories to shop in with current examples. The tooling for building on top of them lives in Part IV. Creative tools get their own deep-dive in Multimodal Creative Tools. As of July 2026.
Coding models
"Best coding model" is now mostly a general-flagship question, but the harness matters as much as the weights (Agent Harnesses & Frameworks), and open models are genuinely competitive.
- Closed leaders: Claude Opus 4.8 / Sonnet 5 (agentic coding, tool use), GPT-5.6 (Codex variant), Claude Fable 5 (premium). Gemini is capable but less commonly reached for by developers for hands-on coding.
- Open leaders: the open-weight models in Open-Weight Models, notably Kimi K3, Qwen3.6-Coder, DeepSeek V4, and GLM-5.2, are genuinely competitive on real-repo tasks (see that table for licenses and sizes).
- Specialized: Cursor's in-house model (tuned for fast in-editor edits), xAI
grok-build(early), and fill-in-the-middle models for autocomplete inside IDE tools. - Judge on SWE-bench Verified, Terminal-Bench, and LMArena's WebDev Arena, not HumanEval (saturated).
Embedding & reranking models
The backbone of RAG and search. Pick on retrieval quality for your domain (run a small eval), then on dimensions (storage/latency cost) and max input length.
| Model | Type | Open? | Notes |
|---|---|---|---|
| OpenAI (text-embedding-3-large) | Embedding | No | Strong default, adjustable dimensions |
| Cohere Embed 4 | Embedding | No | Multilingual, multimodal, long inputs |
| Voyage-3 (voyage-3-large) | Embedding | No | Top retrieval quality, domain variants (code, finance, law) |
| Gemini Embedding | Embedding | No | Google-ecosystem, strong MTEB scores |
| BGE-M3 / GTE-Qwen / Nomic-Embed / E5 | Embedding | ✅ | Best open options; self-host to cut cost |
| Cohere Rerank 3.5 / Voyage Rerank / BGE-reranker / Jina | Reranker | mixed | Big precision boost as a cheap second stage |
Image & video models
Fast-moving, and covered with leaders, licensing, and API notes in Multimodal Creative Tools. In brief: image leaders span Midjourney (aesthetics), FLUX.2 (open weight), Ideogram 3 (text-in-image), Adobe Firefly (commercially indemnified), and native GPT/Gemini generation; video leaders span Sora 2, Veo 3 (native audio), Runway, and Kling, with open Alibaba Wan. Judge on control, licensing, and API access rather than raw "wow," and watch clip length and resolution for video.
Audio & voice models
The model layer under the voice agents in Voice & Real-Time Agents and the music tools in Multimodal Creative Tools.
- TTS: ElevenLabs (quality/voice cloning), Cartesia Sonic (low latency), Deepgram Aura, plus provider TTS (OpenAI, Google).
- STT: Deepgram, AssemblyAI, OpenAI Whisper (open, self-hostable).
- Speech-to-speech / realtime: provider realtime APIs, Amazon Nova Sonic, the basis for voice agents.
- Music: Suno, Udio. → Multimodal Creative Tools
Small & on-device models
For phones, laptops, and edge: Gemma 3n (mobile-optimized), Phi-4-mini, Ministral 3 (3B/8B/14B), Qwen3 0.6-4B, Llama 1B/3B, SmolLM3. Quantize to GGUF/MLX and run via Ollama, llama.cpp, MLX, or LM Studio (Local & Self-Hosting Stack).
Frontier-specialized (research-adjacent)
Worth knowing even if not day-to-day dev APIs: world models (Google's "Omni" / Genie-class, for simulation and embodied agents), science models (AlphaFold-class for biology, materials), and time-series/tabular foundation models. Mostly research or niche-API today, but moving toward productization.