BitCPM CANN 8B — Hardware Requirements & GPU Compatibility
Chat
BitCPM-CANN-8B is OpenBMB's 8-billion-parameter ternary (1.58-bit) quantization-aware variant of its MiniCPM4-8B model, produced by the first publicly reported end-to-end 1.58-bit training system built natively for Huawei's Ascend NPU stack, integrating quantization-aware training into Megatron-LM with MindSpeed acceleration. Against its full-precision MiniCPM4-8B counterpart across 11 benchmarks it retains about 95.7% of performance, and the ternary training approach itself adds only a few percent training overhead on Ascend 910B hardware. This particular checkpoint ships in a "pseudo-quantized" format, with ternary values stored as ordinary floating-point weights, so despite the ternary training it loads and runs exactly like the full-precision 8B model rather than at reduced memory, and needs the same hardware as a dense 8B model to run locally. Context length is 32,768 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in May 2026, alongside smaller 0.5B, 1B, and 3B siblings in the same BitCPM-CANN family.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 32K)Best on 24 GB (Q4_K_M · ~127 tok/s · 32K)Macs from 8 GB (Q2_K · ~19 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run BitCPM CANN 8B
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
ollama run hf.co/betterthings/BitCPM-CANN-8B-Q8_0-GGUFQ4_K_M fits fully on the RTX 4090 · 18.4 GB free · 16K context
Download: BitCPM-CANN-8B-Q8_0-Q4_K_M.gguf4.80 GBbetterthings
Other ways
each opens step-by-stepRun in cloud
Fits on RTX 3060 12GB (6 GB headroom) · Q4_K_M
- Generation speed
- ~45 tok/s
- generation speed
- Cost per 1M output tokens
- $0.51
- per 1M output tokens
How Much VRAM Does BitCPM CANN 8B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 3.8 GB | 4.8 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 32K context · ~174 tok/s | |
| 4.3 GB | 5.3 GB | 3-bit medium — 3-bit medium quantization | Fits: 32K context · ~153 tok/s | |
| 5.2 GB | 6.2 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 32K context · ~127 tok/s | |
| 6.1 GB | 7.1 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 32K context · ~108 tok/s | |
| 7.0 GB | 8 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 32K context · ~94 tok/s | |
| 8.4 GB | 9.4 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 32K context · ~78 tok/s | |
| 16.4 GB | 17.4 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Fits: 32K context · ~40 tok/s |
Selected Q4_K_M
4-bit medium quantization5.6GBat 16K context
- Weights
- 4.8 GB
- KV cache
- 0.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 18.4 GB free · ~127 tok/s
≥50 tok/s needs ≥397 GB/s · RTX 5090 ~225, RTX 3090 Ti ~127
Which Hardware Can Run BitCPM CANN 8B?
Q4_K_M · 5.2 GBBitCPM CANN 8B (Q4_K_M) requires 5.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 7+ GB is recommended. Using the full 33K context window can add up to 1.0 GB, bringing total usage to 6.2 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 151 devices with unified memory can run BitCPM CANN 8B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB).
≤ 11 GB
≤ 11 GB tier →fits · 2.4 GB free · up to 32K ctx
16 GPUs · 1 device family
RTX 3080~96Simulate output on RTX 3080RTX 2080 Ti~77Simulate output on RTX 2080 TiRTX 3070 Ti~77Simulate output on RTX 3070 TiOrin Nano~13Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)20 GB
fits · 14.4 GB free · up to 32K ctx
1 GPU
RX 7900 XT~93Simulate output on RX 7900 XT32–40 GB
fits · 26.4 GB free · up to 32K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~225Simulate output on RTX 5090M5 Max36 GB~62Simulate output on Mac Studio M5 Max (36 GB)M532 GB~21Simulate output on MacBook Air 13" M5 (32 GB)Orin~26Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~17Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| ALoDLM 8BAmazon | — | 7.6 GB | ~119 | same size |
| Llm Jp 4.1 8B Thinkingllm-jp | — | 7.6 GB | ~115 | same size |
| Ling 3.0 TinyInclusion AI | — | 5.5 GB | ~364 | same size |
| LFM2.5 8B A1BLiquid AI | — | 5.6 GB | ~335 | same size |
| Gemma 4 E4BGoogle | — | 5.6 GB | ~123 | same size |
| InternVL3 5 8BOpenGVLab | — | 8.2 GB | ~116 | same size |
Specifications
- Publisher
- OpenBMB
- Family
- BitCPM
- Parameters
- 8B
- Architecture
- MiniCPMForCausalLM
- Context Length
- 32,768 tokens
- Vocabulary Size
- 73,448
- Release Date
- 2026-05-15
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does BitCPM CANN 8B need?
BitCPM CANN 8B requires 5.2 GB of VRAM at Q4_K_M, or 16.4 GB at BF16. Full 33K context adds up to 1.0 GB (6.2 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 8B × 4.8 bits ÷ 8 = 4.8 GB
KV Cache + Overhead ≈ 0.4 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 ≈ 1.4 GB (at full 33K context)
VRAM usage by quantization
What's the best quantization for BitCPM CANN 8B?
For BitCPM CANN 8B, Q4_K_M (5.2 GB) offers the best balance of quality and VRAM usage. Q5_K_M (6.1 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 3.8 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run BitCPM CANN 8B on a Mac?
Yes — Mac Mini M2 (8 GB) and 145 other Macs can run BitCPM CANN 8B, the smallest at Q2_K. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 5.2 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M2 Pro (16 GB).
Can I run BitCPM CANN 8B locally?
Yes — BitCPM CANN 8B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is BitCPM CANN 8B?
At Q4_K_M, BitCPM CANN 8B can reach ~928 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~127 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 ÷ 5.17 × 0.65 = ~1006 tok/s
Estimated speed at Q4_K_M (5.2 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of BitCPM CANN 8B?
At Q4_K_M, the download is about 4.80 GB. The full-precision BF16 version is 16.00 GB. The smallest option (Q2_K) is 3.40 GB.
Which GPUs can run BitCPM CANN 8B?
52 consumer GPUs can run BitCPM CANN 8B at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.
Which devices can run BitCPM CANN 8B?
151 devices with unified memory can run BitCPM CANN 8B at Q4_K_M (5.2 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.