Maxwell team · Released, not on our provider

Maxwell V1 Flash (2026): what it is, who made it, its benchmarks, license and how to run it

Maxwell V1 Flash, named Maxwell-V1.0-Flash on its model card, is an open-weight text-only mixture-of-experts model from the Maxwell team with 35B total and 3B active parameters, a 256K-token context and an Apache 2.0 license, published on Hugging Face in 2026 for reasoning, agentic search and long-context work.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Qwen3.8 Flash, not Maxwell V1 Flash.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
Maxwell team
Type
Text
Released
No date given
OpenRouter id
None

Checked

Maxwell V1 Flash: facts, with sources

  • Hugging Face’s API lists the repository maxwell-model/Maxwell-V1-Flash as created on October 13, 2026 and last modified on November 14, 2026. The model card prints no release date, so we give the repository creation date and do not call it a release date.Source: huggingface.co · checked 2026-11-19
  • The first page of the technical report that the card links (a PDF in the maker’s GitHub repository, read by extracting its text) is dated 2026-10-30 and names the author as Maxwell Team.Source: github.com · checked 2026-11-19
  • GitHub’s API shows the repository maxwell-model/Maxwell-V1-Flash was created on September 28, 2026 and last pushed to on November 13, 2026.Source: api.github.com · checked 2026-11-19
  • The card says the model is "developed by the Maxwell Team" and links a homepage at maxwell-llm.ai, a chat playground and a GitHub repository. Neither the card nor the pages we read name a company or institute behind the team, so we do not name one.Source: huggingface.co · checked 2026-11-19
  • The card describes Maxwell-V1.0-Flash as a text-only causal mixture-of-experts language model with 35B total and 3B active parameters per token (both excluding embeddings) and a configured sequence length of 262,144 tokens, meant for reasoning, agentic and long-context work at a practical inference cost. It has 45 layers (3 dense feed-forward layers and 42 mixture-of-experts layers), a hidden size of 2,048 and a vocabulary of 200,704.Source: huggingface.co · checked 2026-11-19
  • Each mixture-of-experts layer has 256 routed experts, 12 of them active per token, plus 2 shared experts that are always on. The attention uses what the card calls Maxwell Block Sparse Attention: each KV group keeps 8 prefix and 8 local blocks of 128 tokens and adds up to 16 remote blocks picked by a learned indexer, which caps main attention at 4,096 tokens per KV group, and the card says the kernel is 6.7x faster than dense attention at 256K.Source: huggingface.co · checked 2026-11-19
  • The repository’s config.json lists a maximum position length of 524,288, which is double the 262,144 sequence length in the card’s specification table and the 256K the card names elsewhere. We could not reconcile the two and use the card’s figure.Source: huggingface.co · checked 2026-11-19
  • Training, per the card: about 26T tokens of multilingual pre-training with the Muon optimizer, context extended in stages from 4K to 256K, a two-stage supervised fine-tuning, separate reinforcement-learning branches for general tasks, verifiable reasoning and agentic tasks, and a final multi-teacher on-policy distillation that merges them into one model. The tokenizer is described as compact across languages, with low-resource Southeast Asian ones named (Thai, Lao, Myanmar, Khmer, Shan).Source: huggingface.co · checked 2026-11-19
  • The card says Maxwell-V1.0-Flash is released under the Apache License 2.0 and points to the LICENSE file in the repository; Hugging Face’s metadata lists the license as apache-2.0 and the weights are not gated.Source: huggingface.co · checked 2026-11-19
  • Maxwell V1 Flash is not on OpenRouter’s model list (checked November 19, 2026), so this site cannot run it. The card also links a chat playground on the maker’s own site, which we did not test. This site is a text chat with other models, such as Qwen3.8 Flash and DeepSeek V4.1 Flash.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

Maxwell V1 Flash benchmarks: the maker’s own table against four models of similar size

  • Maker’s figures from the card’s post-trained table, not tested by us: BrowseComp 76.3 (82.2 with context management enabled) against 47.0 for Qwen3.6 35B-A3B, HLE with tools 46.1 (55.1), Terminal-Bench 58.65, SWE-Bench Verified 65.4 with the Claude Code harness, and AIME 2026 95.00 on mean@8. The card compares it with four models of comparable size: Qwen3.6 35B-A3B, Gemma-4 26B-A4B-it, Nemotron-3.5 Lightning and Muse Glimmer-30B.Source: huggingface.co · checked 2026-11-19
  • The same table shows weaker rows: IFBench 64.80 against 83.20 for Muse Glimmer-30B, IFEval 93.00 against 96.80, BBEH 52.51 against 64.69, and CoSER role playing 53.54 against 60.32. For the base model the card’s own comparison with two Qwen base models shows C-Eval 77.41 against 88.78 and CMMLU 79.41 against 87.41 for Qwen3.5-w35a3-base, while its long-context RULER score at 256K is 93.35 against 85.75.Source: huggingface.co · checked 2026-11-19
  • How the card says it measured: post-training evaluations mostly use temperature 0.6, top-p 0.95, top-k 20 and a 32K-token context, with 128K for AA-LCR, mathematics and HLE; SWE-Bench Verified runs the Claude Code harness (version 2.1.212) in a network-isolated environment with at most 240 turns and four hours per task; the base-model comparison uses the maker’s own internal evaluation framework.Source: huggingface.co · checked 2026-11-19

Vendor benchmarks

Maxwell V1 Flash: Maxwell-V1.0-Flash against Qwen3.6 35B-A3B, Gemma-4 26B-A4B-it, Nemotron-3.5 Lightning and Muse Glimmer-30B

People who search for Maxwell V1 Flash want to know whether a model with only 3B active parameters can hold up against its size class. The maker’s table says it does well on agentic search and math: its BrowseComp and HLE-with-tools numbers are far above the other four, and on AIME 2026 it is first. It is not first everywhere: Muse Glimmer-30B leads on instruction following (IFBench, IFEval) and on BBEH, and Gemma-4 leads on LiveCodeBench pass@1 by a hair. The table also leaves out the large closed models, so it says nothing about how Maxwell compares with them. For a model you can chat with today, DeepSeek V4.1 Flash and Qwen3.8 run on this site, and Kimi K3 is another agent-oriented model with its own page.

Our table: 15 rows from the post-trained evaluation table on the Maxwell-V1-Flash Hugging Face card, read on November 19, 2026; the scores are the maker’s own and we did not run them, and an asterisk marks the score with context management enabled · checked 2026-11-19
Benchmark (metric)Qwen3.6 35B-A3BGemma-4 26B-A4B-itNemotron-3.5 LightningMuse Glimmer-30BMaxwell V1.0-Flash
IFBench (score)59.6079.1070.8083.2064.80
IFEval (score)93.8094.0091.5096.8093.00
MultiChallenge (pass@1)62.5965.7661.0770.4572.20
AA-LCR (mean@5)71.7065.7060.3079.6075.60
LongBench V2 (score)58.8053.3046.3038.6059.40
MMMLU (score)89.190.987.890.792.0
AIME 2026 (mean@8)91.2587.5087.0894.1795.00
GPQA-Diamond (pass@1)86.0082.3075.4483.5086.52
HLE, 128K context (score)21.4020.6011.4021.8324.79
LiveCodeBench v6 (pass@1)85.3285.9678.3076.8085.66
Terminal-Bench (mean@5)47.4238.4323.1048.5058.65
SWE-Bench Verified, Claude Code (pass@1)63.4039.2042.6047.6065.40
BrowseComp (pass@1)47.0029.2024.2055.6076.30 / 82.20*
HLE with tools (pass@1)28.3031.7016.6034.4046.10 / 55.10*
MuSiQue-300 (mean@8)49.2549.3042.4046.0063.00

Source: huggingface.co

Our table

Maxwell V1 Flash specifications and what it takes to run it

  • The card’s deployment recipe downloads the weights with the Hugging Face CLI and starts vLLM from a Docker image named maxwellmodel/vllm-openai:v0.30.0 with --trust-remote-code, --block-size 128, one GPU (tensor parallel size 1), the qwen3 reasoning parser and the qwen3_coder tool-call parser. It requires an NVIDIA driver of version 580 or newer (CUDA 13.0), or an extra environment flag on older drivers. The card names no GPU model or memory size.Source: huggingface.co · checked 2026-11-19
  • Hugging Face’s API shows 36.8 billion parameters stored in BF16 and repository files that add up to about 74 GB, which is the amount of disk and GPU memory the weights alone take before any context cache. The card recommends temperature 0.6, top-p 0.95 and top-k 20, thinking mode on by default, and passing the full assistant message back, including reasoning_content and tool_calls.Source: huggingface.co · checked 2026-11-19

Spec sheet

Maxwell V1 Flash: Maxwell-V1.0-Flash specifications as printed on its Hugging Face card

This is the small, practical end of the open-weight range. The weight files add up to about 74 GB, and the card’s recipe starts vLLM on a single GPU with tensor parallel size 1, so a high-memory data-center GPU is the realistic target; the card does not say which one, and we have not run it. Compare that with the 645 GB of Sage S2.1 Flash or the 1.65T parameters of the DeepSeek V4 Pro 0813 NVFP4 checkpoint. The table below lists the numbers people ask for first: size, context, experts, license and the software the card tells you to use.

Our table: the Maxwell-V1-Flash Hugging Face card and API, read on November 19, 2026; the card names no GPU model, so memory needs are only what the weight files imply · checked 2026-11-19
ItemValueNote
Parameters (total / active)35B / 3BBoth excluding embeddings; Hugging Face counts 36.8 billion in the files
Context length262,144 tokens (256K)config.json says 524,288; we use the card’s figure
Experts per layer256 routed, 12 active, 2 shared42 mixture-of-experts layers after 3 dense layers
Weight filesBF16, about 74 GB in totalLoaded with trust_remote_code; the repository ships custom model code
Serving softwarevLLM in the Docker image maxwellmodel/vllm-openai:v0.30.0, one GPUNVIDIA driver 580 or newer (CUDA 13.0); --block-size 128 is required in the card’s command
LicenseApache License 2.0LICENSE file in the repository; the weights are not gated

Source: huggingface.coSource: huggingface.coSource: huggingface.co

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to Maxwell V1 Flash you can use now

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

$0.30 / $1.2 (per million tokens)

  • DeepSeek released V4.1-Flash on September 10, 2026.Source: api-docs.deepseek.com · checked 2026-11-11
  • Only the Flash version of V4.1 exists: DeepSeek has published no V4.1 Pro, and the Pro model on its price page is still deepseek-v4-pro (V4-Pro-0813).Source: api-docs.deepseek.com · checked 2026-11-11
  • DeepSeek’s pricing page says the API model deepseek-flash is now served by the V4.1-Flash-1017 snapshot.Source: api-docs.deepseek.com · checked 2026-11-11

FAQ

Maxwell V1 Flash FAQ

What is Maxwell V1 Flash?

Maxwell V1 Flash is the Hugging Face repository name of Maxwell-V1.0-Flash, a text-only mixture-of-experts language model with 35B total and 3B active parameters per token. The card says it is built for strong reasoning, agentic and long-context work at a practical inference cost, using block-sparse attention so that long inputs stay affordable. It is a model you download and serve yourself.

Who made Maxwell V1 Flash and when was it released?

The card says it is developed by the Maxwell Team, with a homepage at maxwell-llm.ai and a GitHub organization named Maxwell; it does not name a company or institute, so we do not either. It prints no release date. The Hugging Face repository was created on October 13, 2026, the linked technical report is dated October 30, and the GitHub repository was last pushed to on November 13.

How big is Maxwell V1 Flash and what does it need to run?

The weight files are about 74 GB in BF16, for 36.8 billion stored parameters. The card’s recipe runs vLLM from a Docker image on a single GPU, needs an NVIDIA driver of version 580 or newer, and names no GPU model. In practice that points to a high-memory data-center GPU, not a typical laptop; we have not run it.

What license does Maxwell V1 Flash use?

The card says it is released under the Apache License 2.0, a permissive license that allows commercial use, modification and redistribution if you keep the license and notices. Read the LICENSE file in the repository for the exact terms. The model needs the repository’s custom code (trust_remote_code), so review that code before running it.

How good is Maxwell V1 Flash?

By the maker’s own table, which we did not run, it scores 76.3 on BrowseComp, 65.4 on SWE-Bench Verified and 95.00 on AIME 2026 (mean@8), ahead of Qwen3.6 35B-A3B on those rows, but behind Muse Glimmer-30B on IFBench (64.80 against 83.20). It compares only with similar-size open models, so the card says nothing about large closed models.

How long a context and which languages does it support?

The card gives a configured sequence length of 262,144 tokens, and its RULER score for the base model at 256K is 93.35. The card says the tokenizer stays compact across languages, including Thai, Lao, Myanmar, Khmer and Shan, and shows multilingual scores such as 92.0 on MMMLU and 91.0 on CEval. Its Chinese knowledge scores for the base model are lower than Qwen’s.

Can I chat with Maxwell V1 Flash here, and what can I use instead?

No. It is not on OpenRouter’s model list, so this site cannot run it. The card links a playground on the maker’s own site, which we did not test. On this site you can chat with DeepSeek V4.1 Flash: a typical reply costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Qwen3.8 is another.

Is this the official Maxwell V1 Flash website?

No. This page is unofficial and not affiliated with or endorsed by the Maxwell team. Facts link to the pages we read.