Maxwell V1 Flash Review: Coding, Languages and Research Tests
Hands-on tests of Maxwell V1 Flash, an open 35B MoE model: it fixed a real bug and handled 67 languages, but failed a game build and web research.
For Engineering and Executives

How does Maxwell V1 Flash actually perform in real tests?
Maxwell V1 Flash performed unevenly in one creator's hands-on testing. It found and fixed a bug in a working full-stack web app with no hints, and it produced a 67-language drink-name list in about 42 seconds. It then failed to build a working HTML game and produced a shallow, flawed web research report. Its published benchmarks say it is strong at agentic search, but the live research test did not back that up. These are single runs by one tester, so treat them as signals, not verdicts.
TL;DR
- Maxwell V1 Flash is an open-weight model with 35 billion total parameters and 3 billion active, released under an Apache 2.0 license with a 256K context window.
- In testing it fixed a single hidden bug in a Flask web app without hints, and the app's signup, map and game features all worked afterward.
- It handled a 67-language task in about 42 seconds, keeping the order and avoiding duplicates, with a few weak spots in low-resource languages.
- It failed a browser game build: the HTML file looked fine in description but had no working buttons and referred to a basket that did not exist.
- Its research report on the Australian stock market took 49 seconds and contained mismatched source labels, a leftover note to itself and numbers that looked a year old.
- The model's thinking trace was messy, with loops and stray Russian and Chinese characters appearing for no reason.
What is Maxwell V1 Flash?
Maxwell V1 Flash is a newly released open model from a lab called Maxwell. The lab says it builds efficient open models for agentic work. "Agentic" means the model plans, uses tools, writes and analyzes code, and handles long tasks that run over many steps. Its stated belief is that cost, not intelligence, is what stops agents from scaling in production. So it aims for the most capability per token, per second and per dollar.
The headline specs:
| Spec | Value |
|---|---|
| Total parameters | 35 billion |
| Active parameters per word | 3 billion |
| Design | Mixture of experts (MoE) |
| License | Apache 2.0 |
| Weights | Open |
| Context window | 256K |
"Open weight" means you can download the model and run it yourself. "Mixture of experts" means the model is split into many small specialist sections, and a router picks only a handful of them for each word. That is how a 35 billion parameter model only does the work of a 3 billion one at any moment. The tester said this is exactly the kind of specification that made Coen's 35 billion parameter model a darling of so many.
Is Maxwell V1 Flash a real original model?
Open model releases on Hugging Face are sometimes fine-tunes of someone else's model with a new name. The tester checked the files before running anything. The configuration lists the model type as its own Maxwell MoE, not Llama or Qwen, and the readme names no base model.
That is not final proof, and the tester said so. It is a strong sign the model was built from scratch. The architecture, as described, is a fairly standard mixture-of-experts design with a few notable features:
- It keeps the whole history of a long text but does not reread everything. It checks the start, the most recent part and a few useful chunks picked by a small helper.
- A router picks a handful of experts per word, while a couple of always-on experts cover the basics.
- A small extra component guesses a few words ahead, which the tester said helps training.
Nothing here is radically new, and that is fine. Another lab shipping a working MoE model is useful on its own.
How did it do at fixing a bug in a real app?
This was the best result. The tester ran the model through an agent harness called Hermes agent. The target was a full-stack Flask app (Flask is a Python web framework) with a live map, environmental dashboard, user signup and login, and a diamond discovery game with levels and a leaderboard. The app had one bug that was corrupting values across the app. The prompt gave a goal and nothing else: no hints, no hand-holding.
The model found the bug and fixed it. After a reset, the numbers looked right, signup worked, and the "explore and dig" action that had been broken now worked. The tester found another diamond in the game to confirm it.
This matters because debugging someone else's code across a database, a back end and a front end is a realistic task. It is closer to real work than writing a fresh function from a clean prompt.
How did it handle 67 languages?
The Maxwell team claims its tokenizer stays compact across languages, especially low-resource Southeast Asian ones such as Thai, Lao, Myanmar, Khmer, Shan, Balinese, Sundanese and Javanese. (A tokenizer breaks text into the pieces a model reads. Compact means fewer pieces per sentence, which saves cost.)
To test this, the tester asked for a drink name grounded in each culture, across 67 languages. The model returned all 67 in the right order with no duplicate names, in about 42 seconds. That is solid instruction-following on a long list. A similar naming task was run on Qwen4 Max across 82 languages.
There were weak spots. For one language, it simply repeated the language's own name. For Southern Thai, it picked the name of a real dish, and the model flagged that itself but left it unchanged. The tester was not sure how many of the other answers were culturally accurate and asked native speakers to check. That caveat matters: speed and format are easy to verify, but cultural accuracy needs a human who speaks the language.
Why did the game build fail?
The tester asked the model to create a realistic cinematic scene of Filipino-style crispy fried chicken as an HTML file. The description the model gave looked good. The result did not work. The page said to click the left basket, but there was no basket. The tester was playing the cook, and no button responded. The page looked plain and the tester called it a clear failure.
The process was also messy. The model's thinking trace, the visible reasoning it writes before answering, looped and ran long. Russian, Chinese and other unrelated characters crept into it. At times the model started building a Flask app instead, which had nothing to do with the request.
This is a common pattern with new models. A model can be good at a narrow, well-defined task (find the bug in this app) and weak at open-ended creative generation (make something interactive from scratch). One result does not predict the other.
Does it deliver on its research benchmark claims?
The published benchmarks say Maxwell sits at or near the top in reasoning, with a proprietary model edging ahead on a couple of tests. The biggest gap shows in agentic search, where Maxwell leads the comparison group and passes the proprietary models. The comparison set named in the video includes Gemma 4, Qwen 3.6 and Haiku.
The tester then ran a live research job: is the Australian stock market being hit by an AI bubble? The goal was to see whether the model would search on its own, weigh both sides, and back claims with sources. Web search was enabled in the agent, and each test started from a clean slate with no memory from earlier runs.
The model started searching, hit some Macquarie and other data sources, and produced a report with a verdict in 49 seconds. That speed was surprising, and it was a warning sign. The report had real problems:
- It left a note to itself inside the report.
- Several source labels did not match the sources' actual content.
- Some numbers were a year old.
The tester concluded it was leaning on its training data instead of doing proper search, and called the result a cookie-cutter answer. That is the opposite of what the benchmark suggests. A benchmark is a controlled test. A live research task has messy sources, and the model has to check its own work.
What does this mean for you?
If you are not choosing models for a living, the lesson is still practical. A strong benchmark chart tells you a model can do well on someone's test set. It does not tell you it will do well on yours.
If you use AI at work:
- Do not switch tools because of a chart. Run your own three or four real tasks first.
- Check sources yourself. A fast research report with a confident verdict is not the same as a verified one. Open the links and check dates.
- For multilingual work, speed and format are the easy part. Have a native speaker review the output.
If you manage a team or buy tools:
- Open models with Apache 2.0 licenses are attractive because you can run them yourself. But the tester needed about 130 GB of VRAM (graphics memory) at full throttle, using vLLM. Lowering the context size reduces that, but this is not laptop territory.
- Cost per token is a real concern for agents. If the model's lower active parameter count translates to cheaper runs, that matters. But cheaper only counts if the output is usable.
A simple test plan:
| Task type | What to test | What to check |
|---|---|---|
| Debugging | A real bug in a small internal tool | Does it fix it without breaking anything else? |
| Long lists | A 50+ item formatting job | Order, duplicates, accuracy |
| Research | A question you already know the answer to | Are sources real, current and matching their labels? |
| Building | A small interactive page | Do the buttons work? |
Run each at least twice. One run can be luck.
Could you build your own model testing scorecard?
Comparing models on your own tasks is easier if the results live in one place instead of scattered notes. A small internal app could hold your test prompts, record each model's output, and let reviewers score accuracy, sources and usability. With Remy, you describe that tracker in plain English and it builds and deploys it as a full-stack app with a database and sign-in, so your team can log results together. Remy's apps can also call nearly 400 AI models, so the same tool could run a prompt across several models side by side. It is a simple alternative to a shared spreadsheet that nobody keeps updated.
Frequently Asked Questions
Is Maxwell V1 Flash open source?
It is open weight under an Apache 2.0 license, which means you can download and run the model and use it commercially under that license's terms. The tester also found no base model listed in the readme and a model type named Maxwell MoE rather than Llama or Qwen, which suggests it is original.
How much hardware does Maxwell V1 Flash need?
In the tester's setup, running at full context with vLLM used about 130 GB of VRAM. The tester said you can reduce that by lowering the context size in the launch command. Exact requirements for other setups were not tested.
Is Maxwell V1 Flash good at coding?
It was mixed. It fixed a hidden bug in a full-stack Flask app without hints. But it failed to produce a working interactive HTML game, with no functioning buttons and a reference to a basket that did not exist. Its reasoning trace also showed loops and stray characters.
Can I trust its research reports?
Not without checking. In one test, the report came back in 49 seconds with mismatched source labels, a leftover note to itself and numbers that looked a year old. The tester suspected it relied on training data instead of fresh search. Verify any claims and sources before using the output.
Does this one test mean the model is bad?
No. These are single runs by one tester, and the model is a first release from a new lab. The results show where it looks strong and where to be careful, and they are a reason to run your own tests.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles

LongCat 2.5 vs Claude Opus 5: Benchmarks and Real Tests
Live Nerf vs Margin Lab: Do AI Models Really Get Worse?

Why Is Nvidia So Dominant in AI Chips? The CUDA Moat Explained
