Apriel 1.6 15B Thinker — Hardware Requirements & GPU Compatibility
VisionFunctionsApriel 1.6 15B Thinker is ServiceNow's 15-billion-parameter multimodal reasoning model, an update from Apriel-1.5-15B-Thinker that processes both text and images. Post-trained using supervised fine-tuning and reinforcement learning, the card reports a score of 57 on the Artificial Analysis index and shows over 30 percent less reasoning token usage compared to version 1.5. The documentation notes it fits on a single GPU, making quantized versions practical for mainstream consumer cards. The context window holds 262,400 tokens. Released under the MIT license, it allows unrestricted commercial and research use. Published in November 2025, it replaces Apriel-1.5-15B-Thinker with a simplified chat template plus four new special tokens designed to simplify parsing tool calls and final responses.
Runs great on a 24 GB GPU
Runs from 12 GB (Q4_K_M · 8K)Best on 24 GB (Q4_K_M · ~67 tok/s · 32K)Macs from 16 GB (Q4_K_M · ~14 tok/s)
MIT
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Apriel 1.6 15B Thinker
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/bartowski/ServiceNow-AI_Apriel-1.6-15b-Thinker-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 10.8 GB free · 16K context
Download: ServiceNow-AI_Apriel-1.6-15b-Thinker-Q4_K_M.gguf8.79 GBbartowski
Ways to runOther ways
each opens step-by-stepRun in cloud
Your own private copy · Q4_K_M · 24 GB GPU
- The easiest way to try this model
- Running in a few minutes, easy setup
- Chat right here or connect any app
$5+ bonus for new users
Runs on RTX 5080 (16 GB) (2 GB headroom) · Q4_K_M
- Generation speed
- ~64 tok/s
- generation speed
- Cost per 1M output tokens
- $0.85
- per 1M output tokens
How Much VRAM Does Apriel 1.6 15B Thinker Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 7.1 GB | 14.7 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 64K context · ~92 tok/s | |
| 7.3 GB | 14.9 GB | 3-bit small — 3-bit small quantization | Fits: 64K context · ~90 tok/s | |
| 8.1 GB | 15.6 GB | 3-bit medium — 3-bit medium quantization | Fits: 64K context · ~81 tok/s | |
| 8.2 GB | 15.8 GB | 4-bit · legacy — 4-bit legacy quantization | Fits: 64K context · ~80 tok/s | |
| 9.7 GB | 17.3 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 32K context · ~67 tok/s | |
| 11.4 GB | 18.9 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 32K context · ~58 tok/s | |
| 13.1 GB | 20.6 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 32K context · ~50 tok/s | |
| 15.7 GB | 23.2 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 32K context · ~42 tok/s |
Selected Q4_K_M
4-bit medium quantization13.2GBat 16K context
- Weights
- 8.9 GB
- KV cache
- 4.0 GB
- Overhead
- 0.3 GB
Which Hardware Can Run Apriel 1.6 15B Thinker?
Q4_K_M · 9.7 GBApriel 1.6 15B Thinker (Q4_K_M) requires 9.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 13+ GB is recommended. Using the full 262K context window can add up to 64.0 GB, bringing total usage to 73.7 GB. 51 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080 12GB. 171 devices with unified memory can run Apriel 1.6 15B Thinker, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), Mac Mini M2 Pro (16 GB).
16–18 GB
16–18 GB tier →fits · 2.8 GB free · up to 16K ctx
22 GPUs · 2 Mac classes · 5 device families
RTX 5080~64Simulate output on RTX 5080RTX 5070 Ti~60Simulate output on RTX 5070 TiRTX 5080 Laptop GPU~60Simulate output on RTX 5080 Laptop GPUM616 GB~11Simulate output on iMac 24" M6 (16 GB)M3 Pro*18 GB~11Simulate output on MacBook Pro 14" M3 Pro (18 GB)iPad~11Simulate output on iPad Pro M5 13" (16 GB)Snapdragon X Plus (X1P-42-100)~9Simulate output on Snapdragon X Plus (8-core) Copilot+ PC20 GB
fits · 6.8 GB free · up to 32K ctx
1 GPU
RX 7900 XT~49Simulate output on RX 7900 XT32–40 GB
fits · 18.8 GB free · up to 64K ctx
1 GPU · 3 Mac classes · 9 device families
RTX 5090~120Simulate output on RTX 5090M5 Max*36 GB~33Simulate output on Mac Studio M5 Max (36 GB)M632 GB~12Simulate output on iMac 24" M6 (32 GB)RTX Spark N1X~20Simulate output on NVIDIA RTX Spark N1X Laptop (32 GB)Orin~14Simulate output on NVIDIA Jetson AGX Orin 32GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Phi 4 ReasoningMicrosoft | — | 12.5 GB | ~69 | same size |
| Phi 3 Medium 128k InstructMicrosoft | — | 12 GB | ~72 | same size |
| Qwen1.5 14BAlibaba | — | 22.2 GB | ~63 | same size |
| Nemotron Terminal 14BNVIDIA | — | 11.9 GB | ~69 | same size |
| Ministral 3 14B Reasoning 2512Mistral AI | — | 12 GB | ~72 | same size |
| Qwen3 14B BaseAlibaba | — | 11.9 GB | ~69 | same size |
Fine-tunes & variants(1)
Specifications
- Publisher
- ServiceNow-AI
- Parameters
- 14.9B
- Architecture
- LlavaForConditionalGeneration
- Context Length
- 262,400 tokens
- Vocabulary Size
- 131,072
- Release Date
- 2025-11-28
- License
- MIT
Frequently Asked Questions
How much VRAM does Apriel 1.6 15B Thinker need?
Apriel 1.6 15B Thinker requires 9.7 GB of VRAM at Q4_K_M, or 30.5 GB at BF16. Full 262K context adds up to 64.0 GB (73.7 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 14.9B × 4.8 bits ÷ 8 = 8.9 GB
KV Cache + Overhead ≈ 0.8 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 64.8 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Apriel 1.6 15B Thinker?
Yes, at Q8_0 (15.7 GB) or lower. Higher quantizations like BF16 (30.5 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Apriel 1.6 15B Thinker?
For Apriel 1.6 15B Thinker, Q4_K_M (9.7 GB) offers the best balance of quality and VRAM usage. Q4_K_L (9.9 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 4.9 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Apriel 1.6 15B Thinker on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 145 other Macs can run Apriel 1.6 15B Thinker. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 9.7 GB of usable unified memory (RAM minus macOS overhead).
Can I run Apriel 1.6 15B Thinker locally?
Yes — Apriel 1.6 15B Thinker can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Apriel 1.6 15B Thinker?
At Q4_K_M, Apriel 1.6 15B Thinker can reach ~494 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~67 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 9.72 × 0.65 = ~535 tok/s
Estimated speed at Q4_K_M (9.7 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Apriel 1.6 15B Thinker?
At Q4_K_M, the download is about 8.92 GB. The full-precision BF16 version is 29.73 GB. The smallest option (IQ2_XXS) is 4.09 GB.
Which GPUs can run Apriel 1.6 15B Thinker?
51 consumer GPUs can run Apriel 1.6 15B Thinker at Q4_K_M. Top options include AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 6900 XT, AMD Radeon RX 6700 XT. 30 GPUs have plenty of headroom for comfortable inference.
Which devices can run Apriel 1.6 15B Thinker?
171 devices with unified memory can run Apriel 1.6 15B Thinker at Q4_K_M (9.7 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.