inclusionAI releases Ling-3.1-flash, a 560B hybrid-linear MoE with 25B active
inclusionAI has released Ling-3.1-flash, a hybrid-linear MoE with 560B total and 25B active parameters and a stated 1M-token context. Its 72 layers split into 63 KDA and 9 Gated MLA layers, so only a small share of layers hold a conventional KV cache.
LLMScale estimates for one user at 8K context: 537.3 GiB at FP8 (4× B200) and 287.8 GiB at INT4 (fits a 512GB M3 Ultra). The card's SGLang recipe uses 8× H200 or 4× GB300. vLLM needs a dedicated branch.
#moe #kv-cache #long-context #sglang
Ling-3.1-flash · 572.12B parameters · 8k context, one user
FP16 KV cache and vLLM's overhead on each device at its defaults. “Fits on”: the smallest device that holds it alone, or how many of the largest it takes. Change the context, users or GPU in the calculator
Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning MoE with 24B parameters, about 2B active per token and 64 experts with top-4 routing. Context grows from 32,768 to 131,072 tokens, on a hybrid of gated short convolutions and GQA. Liquid reports 184 tokens/s decode on an Apple M5 Max.
LLMScale estimates for one user at 8,192 tokens of context: 45.97 GiB in BF16 (A100 80GB), 23.89 GiB in FP8 (RTX 5090) and 13.54 GiB in INT4 (RTX 4060 Ti 16GB).
#moe #on-device #long-context #quantization
LFM2.5-24B-A2B · 23.97B parameters · 8k context, one user
FP16 KV cache and vLLM's overhead on each device at its defaults. “Fits on”: the smallest device that holds it alone, or how many of the largest it takes. Change the context, users or GPU in the calculator
NVIDIA has released Nemotron 3.5 Super VL, an open multimodal MoE model with 120B parameters, 12B active per token. One model covers text, image, document and video input. The checkpoint uses NVFP4: 4-bit weights and activations on Blackwell, 4-bit weights with 16-bit activations on other GPUs.
All 120B weights must fit in memory even though only 12B run per token. NVIDIA says LiLiCorr speculative decoding speeds generation up to 2.3x on DGX Spark.
vLLM has day-0 support for Meta's Muse Spark Open, an open-weight MoE with 816B total and 42B active parameters and a 1M-token context. Weights come in BF16 and NVFP4, plus MXFP4 from AMD.
46 of its 62 layers use a 2048-token sliding window, so KV cache at 1M tokens is just over a quarter of all-full attention. The example serves NVFP4 at TP8 with expert parallelism and a DSpark draft model for speculative decoding.
NVIDIA released Dory1.0-4B, a 4B looped language model: a shared 20-layer block repeats between 8 input and 8 output layers, with interleaved sliding-window and full attention and up to 256K context. Loop count is set at inference; SWE-bench Verified goes from 50.3 at 1 loop to 61.8 at 4.
Per the model card, last-loop KV caching means extra loops add compute but no KV cache over a non-looped model. LLMScale estimates 11.12 GiB in BF16 and 6.60 GiB in INT4 for one user at 8K context.
#looped-model #kv-cache #nvidia #long-context
NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 · 3.98B parameters · 8k context, one user
FP16 KV cache and vLLM's overhead on each device at its defaults. “Fits on”: the smallest device that holds it alone, or how many of the largest it takes. Change the context, users or GPU in the calculator
NVIDIA released Nemotron-Labs-Dory1.0-4B, an open 4B looped model. A shared 20-layer block sits between 8-layer input and output blocks and runs repeatedly. It mixes sliding-window and full attention and supports up to 256K context.
Per-loop caching grows the KV cache with each loop; last-loop mode keeps it at one-loop size. LLMScale estimates 11.12 GiB in BF16 and 8.05 GiB in FP8 for one user at 8K context. vLLM support is still an open PR.
#looped-models #kv-cache #long-context #vllm
NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 · 3.98B parameters · 8k context, one user
FP16 KV cache and vLLM's overhead on each device at its defaults. “Fits on”: the smallest device that holds it alone, or how many of the largest it takes. Change the context, users or GPU in the calculator
LLMScale