Maxwell-V1.0-Flash: Open 35B MoE model edges out Qwen at coding

A developer team called Maxwell has released Maxwell-V1.0-Flash, a language model anyone can download. It has 35 billion parameters, only about 3 billion of which are active per token, and it ships under the Apache 2.0 license. Unlike many rivals, it isn't a retrained Qwen. The team says it built and trained the model from scratch, and its own tests put it slightly ahead of Qwen3.6-35B-A3B on coding and agent tasks.

Model releases Nov 17, 2026 6 min read Source: Google News (TW)
Illustrative image, AI-generated

What's going on

A developer team called Maxwell has released a language model named Maxwell-V1.0-Flash, Taiwan's Liberty Times (自由時報) reports. The weights use the Apache 2.0 license, which is very permissive. Anyone can download the model for free, run it on their own hardware, fine-tune it and use it commercially. The files are on Hugging Face, and the source code and a guide to running the model are on GitHub. If you don't have the hardware, you can try the model in the team's online playground.

The most interesting part is where the model comes from. Many "new" open models are really just fine-tuned versions of Qwen or Llama. Maxwell says it designed this model itself and trained it from scratch, and it has documented the process from pretraining through post-training.

Architecture: lots of experts, little compute

Maxwell-V1.0-Flash has 35 billion parameters, but only about 3 billion of them do any work for each token (a chunk of a word). That's thanks to its Mixture-of-Experts (MoE) architecture. The model is made up of many specialized subnetworks, called "experts", and a router picks which ones each token needs. Think of a big company where only the relevant department handles each request.

The details, according to the developer:

  • 45 transformer layers, 42 of them MoE layers
  • 256 experts per layer, with 12 selected per token
  • 2 shared experts that always run and cover general knowledge
  • a context window of up to 256,000 tokens, the amount of text the model can take in at once. That's several thick books' worth.

For long texts, the team built its own block-sparse attention. Attention is the mechanism a language model uses to work out how every word in a text relates to every other word, and it gets very expensive on long documents. The sparse version only computes selected blocks instead of every word pair. The developer says that in one benchmark at 256K context, this runs 6.7 times faster than conventional dense attention.

How it was trained

Maxwell says it pretrained the model on about 26 trillion tokens of multilingual data, using two techniques:

  • the Muon optimizer, a newer method for updating the weights step by step during training, considered more efficient than the standard Adam optimizer
  • multi-token prediction, where the model learns to predict several upcoming tokens at once instead of just the next one

After that came two rounds of supervised fine-tuning (the model learns from example answers), reinforcement learning (it gets rewarded for good solutions) and multi-teacher distillation, where several stronger "teacher" models pass their skills on to the smaller model. The goal is better reasoning, coding and tool use.

The tokenizer stands out too. That's the component that splits text into tokens, and this one was designed for Southeast Asian languages such as Thai, Lao, Burmese and Khmer. Many models chop these languages into far more pieces than needed, which wastes compute. The Liberty Times points out that all of these claims come from the team itself. Nobody has independently checked the exact data sources or the training process.

Benchmarks: strong at coding, not best everywhere

In the official tests, the model does best on coding and agent tasks:

Benchmark What it measures Maxwell-V1.0-Flash Comparison
SWE-Bench Verified fixing real bugs in software projects 65.4% Qwen3.6-35B-A3B: 63.4%
Terminal-Bench command-line tasks 58.65% ahead of Muse Glimmer and similarly sized models, per the team
BrowseComp web research tasks 76.3% (82.2% with context management) –

It has weak spots, though. On instruction following and longer conversations, Maxwell trails Gemma and Muse Glimmer in some tests. The base model, meaning the version before post-training, scores lower than Qwen3.5's base model on the Chinese knowledge benchmarks CMMLU and C-Eval. All of these numbers come from the developer too. Whether others get the same results under the same conditions is still an open question.

What you need to try it

Only 3 billion active parameters sounds small, but don't be fooled. For the model to run fast, all 35 billion parameters have to sit in memory. The released files come to about 73.6 GB. The team recommends serving the model with vLLM, software that runs language models efficiently on servers, on a single NVIDIA B300 or at least two H100s. That's data center hardware. As the Liberty Times also notes, a regular gaming GPU with 24 GB of VRAM can't load the full model in this form.

Things get interesting for home users once smaller versions show up. Quantization stores the model's numbers at lower precision. At 4 bits, a 35B model would need roughly 22 to 24 GB, plus extra memory for long contexts. The sources don't mention any official quantized versions (such as GGUF for llama.cpp, LM Studio or Ollama). With MoE models this size, you can usually offload some of the experts to regular system RAM. That works on smaller GPUs or on Macs with lots of unified memory, but it's slower. There are no concrete numbers for this yet. If you don't have the hardware, the team's online playground is the easiest way in.

Our take

The class of models with about 35 billion parameters and roughly 3 billion active is crowded, led by Qwen3.6-35B-A3B. A new model that keeps up there, by its own measurements, stands out because it isn't a Qwen offshoot. Its own architecture, its own training recipe and an open license add real variety to the open model market. The tokenizer for Southeast Asian languages could be a practical advantage for users in that region.

Some caveats remain. The benchmark leads are small (65.4% vs. 63.4% on SWE-Bench) and haven't been independently confirmed. By the team's own numbers, the model sometimes falls behind on instruction following and conversation. Open weights also don't mean the training data and process are fully open. The source doesn't say who is behind the Maxwell team. Independent tests and developers' everyday experience will decide whether the model catches on.

Related videos

Maxwell V1.0 Flash Tested Locally: 35B MoE, 3B Active, Apache 2.0 Fahd Mirza · YouTube

Sources

#open-source#moe#coding#language model#apache-2.0
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.