BitNet b1.58 2B4T
Microsoft's 2.4B-parameter native 1-bit LLM: ternary weights trained from scratch on 4 trillion tokens.
About
BitNet b1.58 2B4T is Microsoft's first official native 1-bit large language model, released April 14, 2025. Instead of quantizing a full-precision model after training, it is trained from scratch with ternary weights (-1, 0, +1) using BitLinear layers, so the 2.4B-parameter model trained on 4 trillion tokens fits in a fraction of the memory of a comparable FP16 model. The weights and the bitnet.cpp inference framework are open source under MIT.
The practical pitch is CPU inference: Microsoft reports up to 6.17x speedups on x86 CPUs and 5.07x on ARM versus full-precision baselines, with energy reductions up to 82.2% and 70% respectively, and claims a 100B-scale BitNet model can run on a single CPU at 5-7 tokens per second. The technical report (arXiv:2504.12285) presents evaluations showing performance competitive with full-precision models of similar size, though those are provider-run numbers. The GitHub repository has drawn over 41,000 stars, and Microsoft has since extended the line with 1-bit embedding models.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No independent public benchmark table verified; Microsoft's technical report (arXiv:2504.12285) claims performance competitive with full-precision models of similar size across language, math, coding, and conversational evals (provider-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: official Microsoft release Apr 14, 2025 with MIT-licensed weights and inference code; 41,000+ GitHub stars; technical report with full eval methodology (arXiv:2504.12285). Open weights. Release facts corroborated across the repo, Wikipedia, and press. Strengths: genuine research novelty (native 1-bit training at 4T tokens), strong community traction, permissive license. Docked for: (1) capability-parity claims are provider-reported with no independent leaderboard replication found; (2) deployment is tied to the bespoke bitnet.cpp stack rather than mainstream runtimes.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Official Microsoft release 2025-04-14; 41,000+ GitHub stars on microsoft/BitNet
- · Technical report arXiv:2504.12285 with full eval methodology; 1-bit embedding models added Jul 2026
- · MIT-licensed weights and inference code
- · Capability-parity claims are provider-reported; deployment tied to bitnet.cpp