Pin XiaomiMiMo/MiMo-V2.6-Pro-MOPD to adea8e2c5373181e5a973fa1ecb343cb31af214b with BF16 KV and the bundled seven-token DFlash draft. Target and draft configs, weight indexes, tokenizer config, and chat template match the pinned Pro-RL recipe inputs.
Validation: lil list --models-config .lil/models; lil render for pulsar and kimi-container with dflash and none; lil check for pulsar with dflash and none. All passed. No live launch, generation evaluation, or container runtime preflight was performed. AI assistance was used.
Pin XiaomiMiMo/MiMo-V2.6-Pro-RL to 54b10491b1811c76aa9681a9d0ff872396a4064c. Use the target-resident BF16 DFlash draft with seven speculative tokens and b12x noncausal sliding attention. Qualified through lil on eight SM120 GPUs with CUDA graphs, text, images, and tool calls.
Qualify MiMo-V2.6 Flash at TP4 with b12x output padding
Validated XiaomiMiMo/MiMo-V2.6-Flash-RL revision 3b38d063180c3e4aed9691fdc735f3d10b266ee4 on four SM120 GPUs. The vLLM b12x block-FP8 adapter pads partial output blocks at load time and trims GEMM outputs. Serving retains b12x loading, linears, routed experts, and language attention, with BF16 KV cache and 0.98 memory utilization. CUDA graph capture completed; eight text checks and three image smoke requests passed. The API exposes 1048576 context tokens and four concurrent sequences. Requires the local vLLM MiMo integration and block-FP8 padding implementation.
Pin XiaomiMiMo/MiMo-V2.6-Flash-RL at 3b38d063180c3e4aed9691fdc735f3d10b266ee4. Use b12x loading, language attention, linear and MoE kernels, BF16 KV, and 98% memory utilization. TP2 serving validated on SM120 with vLLM 8aa55670c375a16d4e0ec3eaafbe52ffd4dbc9b9 plus local MiMo integration and b12x c2dc1cf02295b8241fc6a7728be7b6e8c23dda2f. A 301563-token request completed correctly. Local catalog list/render and all lil preflight checks passed.
Configure ModelOpt mixed NVFP4 dense MLPs and MXFP8 attention/head projections with B12X linears and GDN decode, BF16 KV cache, aligned recurrent caching, and three-token MTP. Local preflight passes all nine checks on the pulsar topology using /data/releases/Qwen3.8-27B-QAD-E1. Inference quality is unqualified; the weight repository upload is pending.
The manifest references local-inference-lab/GLM-5.3-NVFP4-Spark, whose model API returns HTTP 404 RepoNotFound. Removing the entry allows catalog loading to continue without resolving the unavailable repository.
Move quantization into precision and cache representation settings into cache without changing their values. Requires a launcher supporting the precision, cache, and loading manifest sections.
Configure DeepSeek V4 Vision with DSpark depth 7 and adaptive verification
Use the DeepSeek V4 0731 B12X, capacity, and compilation policy with probabilistic DSpark drafting and adaptive verification. The profile requires a runtime implementing DeepSeek V4 vision support.
GLM entries gain W4A16 activation forcing and move recurrent-cache settings to the Flash entries; GLM Flash follows its script for sequences, TP2 batching and KV allocation, and the current DFlash2 draft. Qwen follows its scripts for KV dtype, utilization, sequences, and the TP1 profile. DeepSeek-V4-Flash-0731 serves the qualified fixed K5 DSpark profile and drops standard MTP; DeepSeek-V4-Flash is added for the standard checkpoint. The superseded DFlash2-MXFP8 draft entry is removed.
Owned models and DFlash drafts join the DeepSeek entries. Every entry names its weight repository with model; owned entries follow head, external entries with remote code pin a revision.
Add DeepSeek V4 Flash 0731 and Vision-Exp catalog entries
Both entries use the deepseek-v4 family and pin the upstream commit the launcher qualified. Vision-Exp is served as its text model; the vLLM tree registers no V4 vision tower.
