Love this model 🥰
Hey guys!
Thank you for releasing this beutiful model to our community! I think this model deserves more stars than it currently have. It's better than Qwen3.5-122b and on par or better than Laguna S-2.1.
My question is: what is the maxumum context size for it? 131k or 262k?
Model card says Context Training Schedule 8K -> 32K -> 256K Does this means 256k?
Because quantized version from https://huggingface.co/bloomer010/Ling-3.0-flash-GGUF/ when runned in llama.cpp caps context size to 131k
W llama_context: n_ctx_seq (262144) > n_ctx_train (131072) -- possible training context overflow
W srv load_model: the slot context (262144) exceeds the training context of the model (131072) - capping
I srv load_model: initializing, n_slots = 4, n_ctx_slot = 131072, kv_unified = 'true'
Is this quantized version bug? Or the original MAX context size of the model 131k?
Thank you for releasing this beutiful model to our community! I think this model deserves more stars than it currently have. It's better than Qwen3.5-122b and on par or better than Laguna S-2.1.
I absolutely love this model too. It's just amazing in every way and my new main model I'm using for everything. It's just the perfect size to run locally without offloading to RAM. I’m running it at full context (-ctk q8_0 -ctv q8_0) using 2x NVidia RTX 4090 + 2x Intel Arc A770 using i1-IQ4_XS mradermacher quants.
Is this quantized version bug? Or the original MAX context size of the model 131k?
It is probably a quantitation bug as I'm not having that issue on latest llama.cpp with mradermacher quants. I recommend you use mradermacher quants and 256K context will work: https://huggingface.co/mradermacher/Ling-3.0-flash-i1-GGUF