schema = "m64/capsule@1" description = "Qwen3.8-Flash-Next IQ3_S on llama.cpp v1.0.0 (GB10). A local OpenAI-compatible chat endpoint for Qwen3.8-Flash-Next (Qwen's 125B-A6B MoE, ISTA-DASLab's GSQ-RCO IQ3_S GGUF) on stock llama.cpp v1.0.0 built with CUDA 13 for sm_121, 256K context. Reads images, streams, tool calls. Needs no network." kinds = ["chat"] authors = ["dh7"] license = "mit" [hardware] platform = "linux-aarch64" gpu = "NVIDIA GB10" gpu_count = 1 # Unified memory: shard 1 (55 GB) offloaded whole, the PLE n-gram table (shard 2, 29 GB) read lazily through # the page cache, the picture reader (0.9 GB) and a 256K q8_0 KV cache (12 of the 48 layers have one). # Measured on gx10: ~65 GB in use with a 150K-token prompt. vram_gb = 80 driver = ">=580.65" disk_gb = 90 [runtime] sandbox = "container" # No image: the Dockerfile builds llama.cpp v1.0.0 (sm_121) in NVIDIA's PyTorch container and copies the server # onto ubuntu:24.04, both pinned by digest. gpu = true shm_gb = 2 internal_port = 9000 health = "/health" expected_load_seconds = 900 model_name = "qwen3.8-flash-next-iq3_s-llamacpp" user = "65534:65534" # ISTA-DASLab's GSQ-RCO shards and the picture reader, at the same revision as the Strata capsule. [[artifacts]] type = "hf" repo = "ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF" revision = "ed59f92082b1e93c0e96d60a8b11aab089b52f09" include = ["README.md", "IQ3_S/*.gguf", "mmproj-Qwen3.8-Flash-Next-BF16.gguf"] mount = "/weights/qwen38fn"