Maxwell-V1-Flash is a newly trending open MoE language model on Hugging Face — about 35 billion total parameters with three billion active per token — distinct from the unrelated doubleAI “Maxwell” code-optimization brand. We reviewed the maxwell-model/Maxwell-V1-Flash model card on November 17, 2026, when it showed on the order of 100+ community likes and hundreds of downloads.
If you searched “Maxwell” after reading doubleAI’s GPU savings claims, reset your mental model: this Maxwell is an open text MoE with Apache 2.0 licensing and vLLM deployment notes, not a closed compiler service.
TL;DR specs
| Spec | Value (model card) |
|---|---|
| Architecture | MoE causal LM |
| Total / active params | 35B / 3B (excl. embeddings) |
| Context | 256K tokens configured |
| Attention | MBSA block-sparse + dense prefix |
| Experts | 256 routed, 12 active; 2 shared |
| Layers | 45 transformer layers |
| License | Apache 2.0 |
Maxwell Block Sparse Attention (MBSA)
The card claims each KV group keeps 8 prefix + 8 local 128-token blocks and adds up to 16 remote blocks chosen by a learned indexer — capping main attention work while retaining a full-history KV cache. They report the MBSA kernel is 6.7× faster than dense attention beyond the dense prefix in their internal tests (vendor figure — not independently verified here).
Long-context eval tables show RULER scores at 256K still in the low 90s on their base model row — competitive with Qwen3.5-base in the same table, per their numbers.
Training story (high level)
Public text on the card describes:
- ~26T tokens multilingual pretraining with Muon on many matrices
- Block-sparse attention introduced at 128K mid-training after indexer warmup
- SFT + RL branches (math, code, agents) merged via multi-teacher on-policy distillation
Treat parameter totals and token counts as self-reported unless you reproduce from checkpoints.
Benchmark highlights readers care about
Post-trained Maxwell-V1-Flash rows in the card claim:
- AIME 2026 mean@8 around 95.0 versus mid-90s rivals
- SWE-Bench Verified (Claude Code harness) about 65.4% pass@1
- BrowseComp up to 82.2% with context management enabled (footnote on card)
- Long-context averages slightly ahead of Qwen3.6-35B-A3B on their table
Always read footnotes: some search scores include context management asterisks.
Deployment notes
The card requires NVIDIA driver ≥ 580 (CUDA 13) or VLLM_ENABLE_CUDA_COMPATIBILITY=1 in Docker. Download steps:
export MODEL_DIR=$PWD/Maxwell-V1-Flash
hf download maxwell-model/Maxwell-V1-Flash --local-dir "$MODEL_DIR"
Then run the documented vLLM OpenAI-compatible server docker one-liner from the card (adjust GPUs and shm).
Compare hosting economics to Baseten frontier inference guidance and Together agent eval stacks.
Tool calling and templates
The Hugging Face chat_template_jinja exposes <tool_call>, <think>, and multi-turn tool responses — aligned with agent harnesses like Qwen Code and Spillway guard patterns.
Name collision warning
| Name | What it is |
|---|---|
| Maxwell-V1-Flash | Open MoE LM on HF |
| doubleAI Maxwell | Code execution optimizer (our prior post) |
| NVIDIA Maxwell | GPU architecture generation (historical) |
Documentation and bookmarks should use org slug maxwell-model to avoid mixing repos.
Limitations
- VRAM: 35B MoE BF16 is not a laptop model.
- Evals: Vendor tables, not explainx.ai reruns.
- Org transparency: Maxwell Team is newer than tier-one labs — check commit history and quant releases before production bets.
When to pick Maxwell vs Qwen3.6 35B-A3B
Choose Maxwell if you need 256K sparse attention behavior and Apache weights on your vLLM cluster today. Choose Qwen if you need wider hosted API support and corporate support channels. Many teams will download both and run identical agent prompts — same playbook as K2 Horizon fleet posts.
MoE routing and mHC details
The spec table on the card lists 256 routed experts with 12 active per token, plus 2 shared experts per layer, using mHC (multi-head compressed) residual streams with Sinkhorn–Knopp iterations for routing stability. That is a more exotic stack than early Mixtral clones — inference engines must support custom code (custom_code tag on HF).
If your vLLM build lacks Maxwell kernels, expect fallback paths or build failures until the team publishes compatible versions.
Long-context evaluation methodology
Base-model RULER rows show graceful decay from 8K to 256K, with Maxwell-V1.0-Flash (Base) staying above 93 even at 256K in their table while Qwen3.5-base drifts lower. HELMET scores trail RULER — normal for composite long tasks — but still trend upward versus the Qwen row at 128K.
Post-trained models switch evaluation lengths: 128K for math and HLE, 80K for science, 64K for LiveCodeBench v6, per card footnotes. Reproducers must match generation caps or comparisons lie.
Agent and search benchmarks
The card includes Terminal-Bench, SWE-Bench Verified with Claude Code 2.1.212, Tau3-Bench, MCP-Mark, and a suite of search QA sets (BrowseComp, DeepSearchQA, WideSearch). Footnoted context management boosts some search passes — read asterisks before quoting 82% BrowseComp.
These harnesses align Maxwell marketing with agent product wars covered in Together eval posts and Google GTIG exploit agents — different problem, same buyer psychology.
Quantization and Hugging Face siblings
The Hugging Face model card flags quantized variants (hasQuantizations: true in page metadata when we fetched it), but quant artifacts may lag the BF16 drop. Watch the maxwell-model org for GGUF or FP8 variants before planning single-GPU home labs.
Licensing and compliance
Apache 2.0 enables commercial fine-tunes, but downstream benchmark claims in marketing still require reproduction. Legal teams care; ML engineers should too.
Failure modes when serving
| Issue | Mitigation |
|---|---|
| Driver < 580 | Set VLLM_ENABLE_CUDA_COMPATIBILITY=1 per card |
| OOM on 24 GB | Use published quants or tensor parallel |
| Tool template drift | Pin tokenizer revision on HF commit hash |
| Slow MoE | Verify expert parallelism flags in vLLM |
Research lineage: Muon and distillation
Pretraining notes cite Muon optimizer usage on hidden matrices — timely alongside Echelon MoE training research. Multi-teacher on-policy distillation after independent RL branches mirrors industry pipeline stories without naming Maxwell as a product company.
Who should download this week
- Agent startups needing an Apache weights baseline at ~35B MoE scale.
- Inference vendors testing sparse attention kernels against Qwen3.6.
- Students comparing chat templates with tool XML to MiniMind or Qwen Code configs.
Who should wait
Teams needing hosted SLAs, safety classifiers, or EU AI Act paperwork from a major lab should stay on Alibaba Qwen4 API tiers until Maxwell publishes system cards and red-team results.
Hacker News traction
A linked HN item had minimal comments when we checked. Engagement is primarily Hugging Face likes and GitHub technical report links on the card, including the official GitHub repository and its technical report PDF.
Versioning and reproducibility
Pin lastModified dates from the HF card when filing internal model cards. Maxwell-V1-Flash updated in mid-November 2026; silent card edits can change chat templates without a semver tag. Mirror weights to private object storage if your compliance policy forbids runtime pulls from Hugging Face.
Tokenizer size note
The model card lists 200,704 vocabulary entries, much larger than MiniMind’s 6.4k didactic tokenizer. Embedding tables consume meaningful disk even when only 3B parameters activate per forward pass. Budget extra NVMe for shard downloads before overnight pulls.
If your org already standardized on Qwen3.6 tool schemas, budget an afternoon to map Maxwell XML tool blocks into your harness before flipping production traffic. Small template mismatches look like mysterious refusals in logs even when weights load fine.
Related reading
- doubleAI Maxwell claimed vs verified
- Qwen Code harness guide
- Tencent HY4 open weights plans
- Together AI coding agent evals
- Baseten efficient inference
- Hugging Face OpenRouter routing
- Sarvam open-weight India stack
- Echelon MoE training optimizer
Fine-tuning and enterprise guardrails
Apache weights invite LoRA on domain corpora, but Maxwell’s tool template is complex — fine-tunes must preserve XML tool syntax or break agent harnesses. Enterprise buyers should still run internal red-team suites before exposing Maxwell to customer PII, matching discipline we describe for OpenAI CSAM classifier rollouts and Spillway injection guards.
Multilingual table takeaways
The card’s multilingual average lands near 89.6 post-trained, with strong MMMLU and MGSM rows. Teams serving SEA languages should compare against Sarvam India models and Tencent HY4 on their own eval harness, not vendor tables alone.
Roadmap signals to watch on Hugging Face
Watch for: technical report PDF link on the org page, quantized repos, vLLM version pins, and Model index entries tying Base vs Flash vs Instruct variants. Missing pieces mean paper + weights preview, not finished product.
Memory and tensor parallel planning
At ~736 GB total file size on the model card for BF16 shards, plan tensor parallel width before promising latency SLAs. Multi-GPU nodes should reserve headroom for KV cache growth at 256K context — agent sessions blow past 32K quickly when tool outputs recurse.
For cost modeling, compare $/1M tokens against hosted Qwen API tiers and against RightNOW kernel inference pricing if you only need narrow tasks.
Stress test prompts before you adopt
Run your production JSON tool schema, grep-heavy code edit, and long-context RAG prompts side by side against Qwen3.6-35B-A3B on the same GPU count. Maxwell’s win condition is better accuracy at equal cost, not mystique.
Governance and naming in procurement
Procurement spreadsheets should list maxwell-model/Maxwell-V1-Flash explicitly so legal does not confuse the download with doubleAI services or NVIDIA Maxwell GPUs sitting in the same rack row.
Primary sources: maxwell-model/Maxwell-V1-Flash model card, accessed November 17, 2026.
