Unverified50% confidenceBenchmarkExact time
LiquidAI's LFM2.5-24B-A2B uses sparse MoE to activate only 2B per token out of 24B total parameters, extends context from 32K to 131K, runs at 184 tok/s on an Apple M5 Max, and reaches aggregate throughput of up to 21K tok/s on a single NVIDIA B300.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/14/2026
First Seen
Valid until: 1/12/2027
Sources
Related Entities
Related Claims
UnverifiedLFM2.5-24B-A2B achieves a decoding speed of 184 tokens/s on an Apple M5 Max.75% similarVerified以LLaMA-3 70B（80层、64头、每头128维）为例，单个token的KV Cache约占2.6MB（FP16精度），32K上下文的KV Cache总量约达83GB70% similarUnverifiedThe BF16 weights of the 27B model are approximately 54GB (27B × 2 bytes).70% similarUnverifiedLiquid AI d1-omni-600M 参数规模为 587M，Q8_0 GGUF 量化后约 670 MB，llama.cpp 已合并支持，可在树莓派 5 上通过 llama-server 命令运行。70% similarUnverifiedTaking a 70B BF16 model as an example, generating one token requires reading approximately 140 GB of weights, and when fully utilizing the H100's 3.35 TB/s memory bandwidth, the per-token latency is at least 42 milliseconds, with a theoretical TPS upper limit of approximately 24 tokens/second.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026267API
curl https://kongchang.com/api/v1/knowledge/claims/1026267MCP
get_claim(id=1026267)