Skip to main content
MindStudio
Pricing
BlogAbout
My Workspace
Maxwell V1.0 FlashMaxwell MoE modelrun Maxwell locally

Maxwell V1.0 Flash Tested Locally: Does This 35B MoE Model Hold Up?

Hands-on local test of Maxwell V1.0 Flash, a 35B MoE Apache 2.0 model with 3B active params, covering VRAM needs, debugging, and failures.

Edited by Luis Chavez-Mattos, Director of Product RSS
Maxwell V1.0 Flash Tested Locally: Does This 35B MoE Model Hold Up?

What is Maxwell V1.0 Flash?

Maxwell V1.0 Flash is an open weight mixture of experts (MoE) model from a lab called Maxwell, released under an Apache 2.0 license. It has 35 billion total parameters but only activates about 3 billion per token, which keeps inference cheap relative to its size. It ships with a 256K context window and a tokenizer the Maxwell team says stays compact across low resource languages, including several Southeast Asian ones. The model targets agentic work: planning, tool use, coding, and multi step tasks, on the premise that cost, not raw intelligence, is the real bottleneck for deploying agents in production.

On paper, the spec sheet reads like a close cousin of Qwen’s 35B MoE models, which have become a reference point for this size class. Full-throttle local testing, including a full-stack app debugging task, multilingual tokenizer checks, a code generation request, and a live web research job, shows a model that’s genuinely strong in some areas and shaky in others.

TL;DR

  • Maxwell V1.0 Flash is a 35B parameter MoE model with roughly 3B active parameters, Apache 2.0 licensed, and a 256K context window, aimed squarely at agentic and coding workloads.
  • Local hardware demands are steep, with a full-precision local run consuming around 130GB of VRAM, though context length can be trimmed to reduce that footprint when running via VLLM.
  • Debugging performance was a clear win in testing: the model correctly located and fixed a single hidden bug in a full-stack Flask application without any hints, restoring broken functionality end to end.
  • The multilingual tokenizer claim mostly held up, correctly handling a 67-language prompt with no duplicates, though it stumbled on a couple of specific low-resource languages and even flagged its own uncertainty mid-response.
  • Code generation failed outright on a creative front-end task, producing a broken, non-functional HTML page after a messy thinking process littered with unrelated foreign-language tokens.
  • Benchmark claims of leading agentic search were not reflected in a live web research test, where the model produced a flawed report with mismatched source citations and stale, year-old figures instead of fresh search results.
  • The architecture is a standard MoE design with a router selecting a handful of specialist experts per token, always-on baseline experts, and a sparse-attention-style mechanism for long context, nothing architecturally novel but solidly implemented.
The enterprise platform for building software with AI.
Secure, compliant, auditable apps, built to scale.

How much VRAM does Maxwell V1.0 Flash need to run locally?

Running Maxwell V1.0 Flash at full context locally required about 130GB of VRAM in testing, using VLLM as the serving framework. That’s a substantial requirement, putting it out of reach for single consumer GPU setups and firmly into multi-GPU or rented cloud GPU territory. The 256K context window is a major driver of that footprint. Trimming the context size in the VLLM launch command reduces memory consumption, so users without access to high VRAM clusters can still run the model, just with a shorter effective context. For anyone planning to self-host, the practical move is to size the context window to the actual task rather than defaulting to the maximum, since most agentic workflows don’t need the full 256K window anyway.

Is Maxwell V1.0 Flash good at debugging code?

Yes, this is where the model performed best in testing. Given a full-stack Flask application (a fictional diamond monitoring dashboard with a live map, user accounts, and a discovery game) containing one hidden bug that corrupted values across the app, Maxwell was asked to find and fix the issue with no hints. Running through the Hermes agent harness, the model correctly diagnosed the bug and patched it. After the fix, previously broken features, including account signup and the core “explore and dig” gameplay loop, worked correctly. For a model this size running at 3B active parameters, that’s a meaningful result and suggests real utility for agentic coding and debugging tasks rather than just benchmark gaming.

Does the multilingual tokenizer claim hold up?

Largely, yes. Maxwell’s documentation claims its tokenizer stays compact across low-resource languages, including Thai, Lao, Myanmar, Khmer, Shan, Balinese, Sundanese, and Javanese. A test prompt asked the model to return a culturally grounded drink name for each of 67 languages. It returned all 67 in the correct order with no duplicate names, completing the task in about 42 seconds. That’s a solid result for instruction following across a long, specific list. There were weak spots: for one language the model simply repeated the language’s own name instead of producing an actual drink name, and for Southern Thai it picked a dish name rather than a drink, though it flagged the inconsistency itself without correcting it. Those are minor cracks rather than a broken claim, but they show the multilingual coverage isn’t flawless at the edges.

Is Maxwell V1.0 Flash an original model or a fine-tune?

The model’s Hugging Face files list its own architecture type as “Maxwell MoE” rather than referencing Llama or Qwen, and the documentation lists no base model. That’s a reasonably strong signal the model was trained from scratch rather than being a fine-tune of an existing open model wearing a new name, something that’s worth checking given how often “new” open models on Hugging Face turn out to be lightly adapted versions of existing ones. It’s not airtight proof, but combined with the architecture description, it points to Maxwell being a genuine original release.

› Build a claims-review app for 400 adjusters.
already included:✓ HIPAA✓ access control✓ audit log✓ self-hosted
The enterprise platform for building software with AI.

Architecturally, there’s nothing groundbreaking here. It uses a standard MoE setup: a router selects a handful of specialist experts per token while a couple of always-on experts handle baseline processing, which is how a 35B parameter model only activates around 3B at inference time. For long context handling, it uses a sparse-attention-style approach that focuses on the start of the text, the most recent tokens, and a few selected useful chunks rather than rereading the entire context every time. There’s also a lookahead component that predicts a few tokens ahead during training to improve learning. None of this is new relative to other recent MoE releases, but it’s a capable implementation of a known recipe.

Where does Maxwell V1.0 Flash fall short?

Two areas stood out as weak. First, code generation for open-ended creative tasks failed badly. Asked to generate a cinematic, interactive scene of Filipino-style fried chicken as an HTML page, the model’s reasoning process looped, drifted between different app ideas, and was littered with unrelated Russian, Chinese, and other non-English characters with no apparent purpose. The final output was a broken HTML file where none of the interactive elements, including a described “basket” feature, actually worked.

Second, a live web research task meant to test agentic search exposed real gaps. Asked whether the Australian stock market is being hit by an AI bubble, with web search enabled through the Hermes agent, the model returned a report in under a minute, weighing both sides with a stated verdict. But the report contained citation mismatches where source labels didn’t correspond to the content referenced, and some figures appeared to be a year old, suggesting the model leaned on stale training data rather than genuinely synthesizing fresh search results. That directly contradicts the benchmark claims showing Maxwell leading in agentic search tasks against models like Nemotron, Gemma, Qwen, GLM, and even some proprietary models.

Is Maxwell V1.0 Flash worth running?

It depends heavily on the use case. For debugging and tool-use-style agentic coding, it performed well in testing and the Apache 2.0 license plus MoE efficiency make it an attractive option for teams wanting to self-host without per-token API costs. For multilingual tasks involving lower-resource languages, it’s also a strong, if imperfect, performer. But for creative or open-ended code generation, and for tasks requiring reliable live web research with accurate sourcing, it currently falls short of its own benchmark claims. As a first release from a new lab, that mixed picture isn’t unusual, and the gap between benchmark scores and real-world agentic search performance is worth keeping in mind before deploying it for research-heavy workflows.

Frequently Asked Questions

How many parameters does Maxwell V1.0 Flash have?

It has 35 billion total parameters in a mixture of experts architecture, but only about 3 billion are active per token during inference, which keeps compute costs down relative to a dense model of similar total size.

What license is Maxwell V1.0 Flash released under?

It’s released under the Apache 2.0 license, making it open for commercial and research use without the restrictions some other open model licenses impose.

How much VRAM do I need to run Maxwell V1.0 Flash locally?

A full local run via VLLM used around 130GB of VRAM in testing, largely driven by the 256K context window. Reducing the context size in the launch configuration lowers the VRAM requirement for users without access to that much memory.

Can Maxwell V1.0 Flash reliably do web research?

Testing showed mixed results. It completed research tasks quickly but produced a report with mismatched source citations and some outdated figures, suggesting it sometimes falls back on training data rather than fully grounding answers in live search results.

Let every team build.
Without shadow IT.
IT adminEvery app, approved
Campaign trackerMarketing✓
ExpensesFinance✓
OnboardingHR✓
InspectionsOperations✓
Shift schedulerNursing✓

Is Maxwell V1.0 Flash a fine-tune of another model?

Available evidence, including its model type listed as “Maxwell MoE” and no base model specified in its documentation, suggests it’s an original model built from scratch rather than a fine-tune of Llama, Qwen, or another existing open model.

Editorial standards

Presented by MindStudio

No spam. Unsubscribe anytime.