What is BitNet b1.58 2B4T?
BitNet b1.58 2B4T is a research model from Microsoft Research, released in April 2025. It has about 2 billion parameters, was trained on 4 trillion tokens, and every weight is one of three values: -1, 0 or +1, about 1.58 bits each. Activations are 8-bit. Unlike most small quantised models, it was trained this way from the start rather than compressed after training.
The point is efficiency. Microsoft's table puts its non-embedding memory at 0.4 GB, against 1.4 GB for Gemma 3 1B and 2.6 GB for Qwen2.5 1.5B, with 29 ms per token when decoding on a CPU and an estimated 0.028 joules per token. It reads and writes English text, with a 4,096-token context and the Llama 3 tokenizer.
It is under the MIT licence. Those gains need Microsoft's bitnet.cpp runtime; in standard transformers it runs no faster than a full-precision model. The bitnet.cpp repository has about 41,500 GitHub stars and was last updated in July 2026; the weights drew about 11,000 Hugging Face downloads in the last 30 days.
Is BitNet b1.58 2B4T worth it?
BitNet's result is that ternary weights trained from scratch hold up. Across Microsoft's benchmark set it averages 54.19, against 55.23 for Qwen2.5 1.5B, 48.70 for SmolLM2 1.7B, 44.90 for Llama 3.2 1B and 43.74 for Gemma 3 1B. It leads that group on GSM8K maths at 58.38 and on ARC-Challenge at 49.91, while using 0.4 GB of non-embedding memory against 2.6 GB for Qwen.
It is weaker where knowledge and code matter. Qwen2.5 1.5B beats it on MMLU (60.25 to 53.17), HumanEval+ (50.60 to 38.40) and MATH-500 (53.00 to 43.40), and Gemma 3 1B leads on instruction following. Those rivals are themselves 2024 and early-2025 models, so the table says little about how it compares with the small models released since.
The practical catch is the runtime and the context. The speed and energy savings only appear in bitnet.cpp; Microsoft warns that transformers runs it as slowly as, or slower than, a normal model. The 4,096-token context is short for chat history or documents, support for languages other than English is limited, and Microsoft's own card says it does not recommend commercial or real-world use without further testing.
The value is as a reference point. It is the model to study if you are deciding whether 1-bit inference on CPUs is worth building for, and Microsoft has not yet released a larger or newer BitNet language model to replace it.
Where it earns its place
An embedded software engineer at an industrial sensor maker
Test whether a language model can run on a gateway with no GPU and little RAM
Build bitnet.cpp and run the GGUF on the target CPU. Microsoft measured 29 ms per token and 0.4 GB of non-embedding memory; check the 4,096-token context fits your prompts.
A PhD student in an efficient-ML group
Fine-tune a ternary model for a narrow task and compare it with a quantised baseline
Start from the BF16 master weights, which Microsoft publishes for training, under the MIT licence. The 4 trillion training tokens and published recipe give a clean baseline against post-training quantisation.
Before you commit
- Setup
- Clone and compile microsoft/BitNet (bitnet.cpp), download the GGUF and run it from the command line. The transformers route needs a specific fork and gives no speed benefit.
- Integrations
- bitnet.cpp for CPU inference, a transformers fork for experiments, and BF16 weights for training or fine-tuning.
- Data and privacy
- Runs fully offline on your own CPU. MIT licence allows commercial use, though Microsoft recommends further testing first.
- Cost at scale
- Weights are free and run on ordinary CPUs; Microsoft estimates 0.028 J per token against 0.186 J for Gemma 3 1B.
- Limits
- 4,096-token context, limited non-English support, and an elevated error rate on election questions, per Microsoft's card.
What’s good
- 0.4 GB non-embedding memory, about a sixth of Qwen2.5 1.5B
- 29 ms per token on CPU in Microsoft's test
- Near Qwen2.5 1.5B on average (54.19 to 55.23)
- MIT licence, with BF16 master weights for fine-tuning
What’s not
- Savings need bitnet.cpp; transformers is no faster
- 4,096-token context
- Microsoft advises against commercial use without testing
- Compared only against 2024 and early-2025 models
Best for
Researchers and embedded engineers evaluating whether native 1-bit models are viable on CPU-only hardware.
BitNet b1.58 2B4T alternatives
BitNet b1.58 2B4T: common questions
- Is BitNet b1.58 2B4T worth it?
- BitNet b1.58 2B4T proves a native 1-bit model can match 2B full-precision rivals on a fraction of the memory, but it is a research artefact, not a model to ship. We rate it 3.5 out of 5.
- How much does BitNet b1.58 2B4T cost?
- Free MIT-licensed weights and bitnet.cpp runtime. No hosted API.
- What are the downsides of BitNet b1.58 2B4T?
- Savings need bitnet.cpp; transformers is no faster; 4,096-token context; microsoft advises against commercial use without testing; compared only against 2024 and early-2025 models.
- Who is BitNet b1.58 2B4T best for?
- Researchers and embedded engineers evaluating whether native 1-bit models are viable on CPU-only hardware.
- How much does BitNet b1.58 2B4T cost for a team or in production?
- Weights are free and run on ordinary CPUs; Microsoft estimates 0.028 J per token against 0.186 J for Gemma 3 1B.
- Is BitNet b1.58 2B4T safe for company data?
- Runs fully offline on your own CPU. MIT licence allows commercial use, though Microsoft recommends further testing first.
- What are the best BitNet b1.58 2B4T alternatives?
- The closest alternatives we have reviewed are Ternary Bonsai 2 9B, SmolLM3-3B, MiniCPM5-2B and LFM3.
Answers checked Nov 7, 2026.