Maxwell V1 Flash: An MoE Model for Inference and Agents with Low Activation Volume

maxwell-model: An MoE language model (internal name Maxwell-V1.0-Flash) created on 10 13 and updated on 11 14, with 35B total parameters, 3B active parameters per token (excluding embeddings), a 256K context window, and text-only modality. It uses proprietary Block Sparse Attention MBSA (limiting primary attention per KV group to 4,096 tokens; officially claimed to be 6.7x faster than dense GQA kernels at 256K context), 4-path residual stream mHC, and fine-grained experts (selecting 12 out of 256, plus 2 shared experts); pre-trained on approximately 26T tokens in multiple languages, with a tokenizer that is more efficient for Southeast Asian languages such as Thai, Lao, Burmese, and Khmer. Licensed under Apache-2.0.

Original title: maxwell-model/Maxwell-V1-Flash

maxwell-model: An MoE language model (internal name Maxwell-V1.0-Flash) created on 10 13 and updated on 11 14, with 35B total parameters, 3B active parameters per token (excluding embeddings), a 256K context window, and text-only modality. It uses proprietary Block Sparse Attention MBSA (limiting primary attention per KV group to 4,096 tokens; officially claimed to be 6.7x faster than dense GQA kernels at 256K context), 4-path residual stream mHC, and fine-grained experts (selecting 12 out of 256, plus 2 shared experts); pre-trained on approximately 26T tokens in multiple languages, with a tokenizer that is more efficient for Southeast Asian languages such as Thai, Lao, Burmese, and Khmer. Licensed under Apache-2.0.

Why it matters Activation volume is comparable to Qwen models in the same tier; officially claimed higher scores on search and terminal tasks; official examples show single-GPU deployment; Apache license

Rising score 154 · Likes 156 · Downloads 287

Reasons

  • Official comparisons against 4 peer models including Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it: BrowseComp 76.30 (with context management 82.20), Terminal-Bench 58.65, SWE-Bench-Verified (Claude Code framework) 65.40, and AIME 2025 mean@8 96.25—all the highest in the table
  • There are also lagging metrics: IFBench 64.80, BBEH 52.51, and CoSER 53.54 are lower than several models in the table; the average instruction-following score is lower than Gemma 4 and Muse Glimmer
  • Official resources include a customized vLLM image, a technical report, and an online demo page

Usage

  • Use hf download maxwell-model/Maxwell-V1-Flash to download locally, then run vllm serve using the Docker image maxwellmodel/vllm-openai:v0.30.0 with flags --trust-remote-code --block-size 128 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3\_coder
  • Thinking mode is enabled by default; for multi-turn conversations and tool calls, pass back the previous turn's reasoning\_content and tool\_calls exactly as they were; official evaluations typically use temperature 0.6, top_p 0.95, and top_k 20
  • Requires NVIDIA driver version 580 or higher (CUDA 13.0); for older drivers, add -e VLLM\_ENABLE\_CUDA\_COMPATIBILITY=1 in the docker run command

Risks

  • The model card does not introduce the development team or institutional background, providing only a contact email
  • Relies on the official customized vLLM image and trust\_remote\_code; the official documentation does not specify VRAM requirements, and loading the full 35B weights on a single GPU requires substantial VRAM
  • The repository was created on 10 13 but only recently gained traction; all scores come from the official internal evaluation framework. This entry has not been independently run

Open

Hugging Face Trending

Explore the latest AI Radar updates