AI

Maxwell-V1.0-Flash: An Open AI Model With 74 GB of Weights

Maxwell-V1.0-Flash: 35B parameters, Apache 2.0, 73.6 GB of weights. What it takes to run it and where it beats and trails its rivals.

The Maxwell team has published an open language model, Maxwell-V1.0-Flash, on Hugging Face: 35B parameters in total, 3B of them active per token, with a context of up to 256,000 tokens. It is a text-only mixture-of-experts model with 256 routed experts per layer, of which 12 work on each token alongside 2 shared ones. The license is Apache 2.0 and the weights can be downloaded without requesting access. The developers describe the goal as strong reasoning, agentic work and long context at a practical inference cost.

The weights take 73.6 GB in BF16. For deployment the developers provide a ready Docker image with vLLM (maxwellmodel/vllm-openai v0.30.0) that starts an OpenAI-compatible server. It needs an NVIDIA driver version 580 or newer (CUDA 13.0), the example launch uses a single GPU, and the model loads with custom code (the trust-remote-code flag). Thinking mode is on by default and function calling is supported. There is also an online playground at maxwell-llm.ai/playground.

For long texts the model uses block-sparse attention, MBSA: each key group looks at blocks from the start of the text, the nearest 128-token blocks and up to 16 remote blocks picked by a learned indexer. Main attention is bounded at 4,096 tokens while the full history stays in the cache. According to the developers, the MBSA kernel is 6.7 times faster than ordinary dense attention at 256,000 tokens. Pre-training used about 26T tokens of multilingual data, and the tokenizer is described as compact for Thai, Lao, Myanmar, Khmer and Shan.

The results are strongest where the model works as an agent. On SWE-Bench Verified (the share of real software bugs fixed, run with Claude Code) Maxwell scores 65.4 against 63.4 for Qwen3.6-35B-A3B, 47.6 for Muse-Glimmer-30B and 39.2 for Gemma-4-26B-A4B-it. On Terminal-Bench (command-line tasks) it gets 58.65 against 48.50 for Muse-Glimmer-30B. On BrowseComp (finding answers on the web) it reaches 76.3, or 82.2 with context management, while Muse-Glimmer-30B gets 55.6 and Qwen3.6 gets 47.0. Nemotron-3.5-Lightning is also in the comparison.

The model is weaker at following instructions precisely: its average score is 76.37 against 84.60 for Muse-Glimmer-30B and 83.07 for Gemma 4, and on IFBench it gets 64.8 against 83.2 for Muse-Glimmer-30B. Muse-Glimmer-30B also leads on multi-turn dialogue (81.46 on average against 79.36) and on logic tasks (78.03 against 73.58). The base version, before post-training, scores 79.41 and 77.41 on the Chinese tests CMMLU and C-Eval, against 86.17 and 86.33 for Qwen3.5-w35a3-base. All measurements were made by the Maxwell team using its own methodology.

The 35B-total, 3B-active class is the same as Qwen3.6-35B-A3B in the developers’ table. There are no compressed (quantized) versions of Maxwell on the page yet, and the repository, created on October 13, 2026, shows 154 downloads and 43 likes. KRAKOBIT has covered what hardware local models need in its piece on a local 27B AI model, and running a large model on a single RTX 3090 with 64 GB of RAM in the news on DeepSeek V4.1 Flash.