What is Maxwell-V1.0-Flash?
Maxwell-V1.0-Flash is a text-only mixture-of-experts language model with 35 billion total and 3 billion active parameters per token (both excluding embeddings) and a 256K-token context window. It has 45 layers: three dense, then 42 with 256 routed experts, of which 12 are active per token, plus 2 shared experts. Attention uses Maxwell Block Sparse Attention, which keeps 8 prefix and 8 local 128-token blocks and picks up to 16 more with a learned indexer, capping main attention at 4,096 tokens per KV group; the card says the kernel is 6.7x faster than dense attention at 256K.
It was pre-trained on about 26 trillion multilingual tokens with the Muon optimizer, with a tokenizer built to stay compact on Thai, Lao, Myanmar, Khmer and Shan, then post-trained with supervised fine-tuning, separate reinforcement-learning branches for general, reasoning and agentic tasks, and multi-teacher distillation into one model. Weights are Apache 2.0. The publisher is the "Maxwell Team"; the model card, technical report and website name no company, location or funder.
Is Maxwell-V1.0-Flash worth it?
The headline results are agentic. In the post-training table Maxwell scores 76.30 on BrowseComp against 47.00 for Qwen3.6-35B-A3B and 55.60 for Muse Glimmer-30B, 83.30 on DeepSearchQA, and 58.65 on Terminal-Bench against 47.42 for Qwen3.6. On SWE-Bench Verified, run in a Claude Code harness, it edges Qwen3.6 by 65.40 to 63.40. Its six-benchmark mathematics average is 92.61 (mean@8), and it leads the long-context average at 76.94.
It is weaker at doing what it is told. IFBench is 64.80 against 83.20 for Muse Glimmer and 79.10 for Gemma 4 26B-A4B, and the instruction-following average trails both. The base model also scores below the Qwen base models on Chinese (CMMLU 79.41 against 87.41), though it leads three of the four Southeast Asian base benchmarks, where the tokenizer work shows.
The catch: every score comes from the team's internal evaluation, with settings it chose (temperature 0.6, contexts from 32K to 256K, and context management enabled for the starred search results), and there is no independent leaderboard entry yet. The organisation behind it is unidentified: a 76-name author list, a contact email and a website, but no company, country or funding. Running it needs trust_remote_code, which executes the repository's Python on your machine, and a custom vLLM image built for NVIDIA driver 580 or newer. The safetensors hold about 36.8 billion BF16 parameters, roughly 74 GB, before any KV cache. The card also lists a RoPE base of 10,000,000 where config.json says 20,000,000, a small sign of a rushed release.
Uptake is modest: 215 likes and 1,052 downloads on Hugging Face, with a third-party GGUF build at 4,437. The weights are Apache 2.0, so the licence is not the obstacle; provenance and verification are.
Our take: Maxwell earns a trial from anyone running search or research agents on their own GPUs, measured on their own tasks against Qwen3.6-35B-A3B; for anything regulated, an unnamed maker and remote code are reasons to wait for outside benchmarks.
Where it earns its place
An ML engineer at a market-research firm
Run a deep-research agent that browses and cites sources on our own hardware
Serve the model with the maxwellmodel/vllm-openai image and the qwen3 tool-call parser, then plug it into the existing agent through the OpenAI-compatible API. Compare it with Qwen3.6-35B-A3B on a set of your past research questions before switching.
A developer building a support assistant for Thai and Vietnamese customers
Get good answers in Southeast Asian languages without paying for long prompts
The tokenizer is the most compact in the technical report's table on Southeast Asian text, so the same prompt uses fewer tokens and less context. Check output quality on your own transcripts, since the multilingual scores are self-reported.
Before you commit
- Setup
- Download the weights with hf download, then run vLLM from the maxwellmodel/vllm-openai:v0.30.0 Docker image with --trust-remote-code; needs NVIDIA driver 580 or newer (CUDA 13.0), or a compatibility flag on older drivers.
- Integrations
- OpenAI-compatible chat API through vLLM with tool calling and a thinking mode on by default; a community GGUF build exists on Hugging Face.
- Data and privacy
- Runs on your own hardware under Apache 2.0. Loading it executes the repository's custom model code. The web playground's data handling is not published.
- Cost at scale
- Free weights; about 74 GB of BF16 weights before KV cache, so one large data-centre GPU or several smaller ones. The README example uses a tensor-parallel size of 1.
- Limits
- Text only; benchmarks are internal and partly use context management; maker's identity, location and training compute are not published.
What’s good
- Leads its comparison table on BrowseComp, DeepSearchQA and Terminal-Bench
- 3B active parameters with block-sparse attention to 256K
- Compact tokenizer for Southeast Asian languages
- Apache 2.0 weights with a 38-page technical report
What’s not
- All benchmarks are self-reported
- Publisher's company and country are not published
- Needs trust_remote_code and a custom vLLM image
- Instruction following trails Gemma 4 and Muse Glimmer
Best for
Teams building search and research agents who can host a roughly 74 GB model and want to test a 3B-active alternative to Qwen3.6-35B-A3B.
Maxwell-V1.0-Flash alternatives
Maxwell-V1.0-Flash: common questions
- Is Maxwell-V1.0-Flash worth it?
- On its makers' own tests Maxwell is the strongest 3B-active model for agentic search, but every number is self-reported and the company behind it is not named. We rate it 3.7 out of 5.
- How much does Maxwell-V1.0-Flash cost?
- Free open weights under Apache 2.0. A web playground is offered; no paid API or pricing is published.
- What are the downsides of Maxwell-V1.0-Flash?
- All benchmarks are self-reported; publisher's company and country are not published; needs trust_remote_code and a custom vLLM image; instruction following trails Gemma 4 and Muse Glimmer.
- Who is Maxwell-V1.0-Flash best for?
- Teams building search and research agents who can host a roughly 74 GB model and want to test a 3B-active alternative to Qwen3.6-35B-A3B.
- How much does Maxwell-V1.0-Flash cost for a team or in production?
- Free weights; about 74 GB of BF16 weights before KV cache, so one large data-centre GPU or several smaller ones. The README example uses a tensor-parallel size of 1.
- Is Maxwell-V1.0-Flash safe for company data?
- Runs on your own hardware under Apache 2.0. Loading it executes the repository's custom model code. The web playground's data handling is not published.
- What are the best Maxwell-V1.0-Flash alternatives?
- The closest alternatives we have reviewed are Qwen3.6, Muse Glimmer and Gemma 4.
Answers checked Nov 18, 2026.