New market map: on-device models in 2026 — 119 companies, 14 layers The valuations sit with the sma...
New market map: on-device models in 2026 — 119 companies, 14 layers
The valuations sit with the small-model makers, not the chips. Mistral is valued at over EUR 21B and Liquid AI at $2B+, while Ollama, a one-line local runner, has 185K GitHub stars and llama.cpp 133K.
Layer by layer:
1. Big-tech small models (17) — Mistral Ministral (EUR 21B+ val.), Qwen small models, Google Gemma / Gemini Nano, Microsoft Phi, IBM Granite, Meta Llama 3.2 1B/3B.
2. Startup small models (10) — ModelBest (MiniCPM, 24M+ downloads), Multiverse ($1.7B val.), Liquid AI ($2B+), Sarvam ($1.5B), Arcee, Zyphra.
3. Speech recognition (13) — Deepgram ($1.3B val.), Whisper (5.6M Hugging Face downloads in 30 days), Argmax, NVIDIA Parakeet.
4. Speech synthesis (11) — Sesame ($250M raised), Cartesia ($100M), Fish Audio ($21M ARR), Kokoro (11.5M downloads in 30 days).
5. Full-duplex, wake word & in-car voice (8) — Cerence ($251.8M revenue), SoundHound ($61.9M quarterly revenue), Kyutai, Sensory.
6. Local runners (5) — Ollama (185K stars), GPT4All (77K), Jan (45K), LM Studio.
7. Inference engines (11) — llama.cpp (133K stars), ncnn (24K), MLC LLM (23K), ONNX Runtime (22K), KTransformers.
8. Chip & platform toolchains (8) — Qualcomm AI Hub (Snapdragon X2: up to 80 TOPS), Apple Core ML / MLX, Google LiteRT, Intel OpenVINO.
9. TinyML (2) — Arduino (33M+ developers), Edge Impulse.
10. Phone & PC NPUs (8) — Qualcomm Hexagon, AMD XDNA (60 TOPS), Intel (50 TOPS), MediaTek, Apple Neural Engine, Huawei Kirin.
11. Standalone edge NPUs (13) — SiMa.ai ($1.45B val.), Axelera AI ($250M+ raised), Ambarella ($390.7M revenue), Hailo (40 TOPS), DeepX, Syntiant.
12. NPU IP (3) — Arm Ethos, Expedera ($20M), Quadric (up to 864 TOPS).
13. System-level AI from US platforms (4) — Apple Intelligence, Google AICore, Microsoft Foundry Local, Samsung Galaxy AI.
14. China phone makers (6) — OPPO, vivo, Xiaomi, Huawei, Honor and Nubia, each with an on-device model.
Data as of Nov 2026, public sources; company-reported numbers are labeled as such on the map.
Bookmark this — we'll update it every quarter.
Full list + sources: edge-ai-models-map.vercel.app
The valuations sit with the small-model makers, not the chips. Mistral is valued at over EUR 21B and Liquid AI at $2B+, while Ollama, a one-line local runner, has 185K GitHub stars and llama.cpp 133K.
Layer by layer:
1. Big-tech small models (17) — Mistral Ministral (EUR 21B+ val.), Qwen small models, Google Gemma / Gemini Nano, Microsoft Phi, IBM Granite, Meta Llama 3.2 1B/3B.
2. Startup small models (10) — ModelBest (MiniCPM, 24M+ downloads), Multiverse ($1.7B val.), Liquid AI ($2B+), Sarvam ($1.5B), Arcee, Zyphra.
3. Speech recognition (13) — Deepgram ($1.3B val.), Whisper (5.6M Hugging Face downloads in 30 days), Argmax, NVIDIA Parakeet.
4. Speech synthesis (11) — Sesame ($250M raised), Cartesia ($100M), Fish Audio ($21M ARR), Kokoro (11.5M downloads in 30 days).
5. Full-duplex, wake word & in-car voice (8) — Cerence ($251.8M revenue), SoundHound ($61.9M quarterly revenue), Kyutai, Sensory.
6. Local runners (5) — Ollama (185K stars), GPT4All (77K), Jan (45K), LM Studio.
7. Inference engines (11) — llama.cpp (133K stars), ncnn (24K), MLC LLM (23K), ONNX Runtime (22K), KTransformers.
8. Chip & platform toolchains (8) — Qualcomm AI Hub (Snapdragon X2: up to 80 TOPS), Apple Core ML / MLX, Google LiteRT, Intel OpenVINO.
9. TinyML (2) — Arduino (33M+ developers), Edge Impulse.
10. Phone & PC NPUs (8) — Qualcomm Hexagon, AMD XDNA (60 TOPS), Intel (50 TOPS), MediaTek, Apple Neural Engine, Huawei Kirin.
11. Standalone edge NPUs (13) — SiMa.ai ($1.45B val.), Axelera AI ($250M+ raised), Ambarella ($390.7M revenue), Hailo (40 TOPS), DeepX, Syntiant.
12. NPU IP (3) — Arm Ethos, Expedera ($20M), Quadric (up to 864 TOPS).
13. System-level AI from US platforms (4) — Apple Intelligence, Google AICore, Microsoft Foundry Local, Samsung Galaxy AI.
14. China phone makers (6) — OPPO, vivo, Xiaomi, Huawei, Honor and Nubia, each with an on-device model.
Data as of Nov 2026, public sources; company-reported numbers are labeled as such on the map.
Bookmark this — we'll update it every quarter.
Full list + sources: edge-ai-models-map.vercel.app
💬1🔄0❤️0👀638📊1