Maxwell-V1-Flash-GGUF 35B MoE Model: 24GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | AtomicChat/Maxwell-V1-Flash-GGUF |
| Published | 2026-11-17 |
| License | apache-2.0 |
| Formats | GGUF |
| Source type | Unverified (not confirmed by a primary source) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
This information has not been officially confirmed. This report is based on the description in the AtomicChat repository AtomicChat/Maxwell-V1-Flash-GGUF on Hugging Face.
It is reported to be a GGUF conversion of the Maxwell team’s Maxwell-V1-Flash by the Atomic Chat team (AtomicChat). The team explains that they wrote a custom support patch for llama.cpp themselves. However, it is noted that a llama.cpp patched with this is required, and it cannot yet be loaded with standard llama.cpp, LM Studio, or Ollama.
Specifications
Below are the details from the original model card and the GGUF version card.
- Parameters: Total 35B, 3B active per token (both excluding embeddings)
- Architecture: MoE (text-only causal LM). 45 layers (3 Dense FFN layers + 42 MoE layers), with 256 routed experts and 12 active
- Attention: GQA (32 query heads / 4 KV heads). The original model uses block-sparse attention
- Context length: 256K (stated in card. Config on the GGUF side is 524,288)
- Vocabulary size: 200,704
Performance
Below are the publisher’s own measurements found on the card for the original model maxwell-model/Maxwell-V1-Flash (post-trained). These are not verified by a third party. The table restricts comparison targets to four columns and selectively trims rows due to space constraints. The table in the card does not run all the way through, and agent-related rows could not be confirmed.
→ Scroll horizontally to see all columns
| Benchmark | Qwen3.6 35B-A3B | Gemma-4 26B-A4B-it | Muse Glimmer-30B | Maxwell V1.0-Flash |
|---|---|---|---|---|
| Instruction following | ||||
| IFEval | 93.80 | 94.00 | 96.80 | 93.00 |
| Multi-turn | ||||
| MultiChallenge (pass@1) | 62.59 | 65.76 | 70.45 | 72.20 |
| Long context | ||||
| AA-LCR (mean@5) | 71.70 | 65.70 | 79.60 | 75.60 |
| LongBench V2 | 58.80 | 53.30 | 38.60 | 59.40 |
| RULER | 97.08 | 94.74 | 94.84 | 95.81 |
| Multilingual | ||||
| MGSM | 90.3 | 91.1 | 93.2 | 93.0 |
| Mathematics | ||||
| AIME 2025 (mean@8) | 90.00 | 85.83 | 92.08 | 96.25 |
| HMMT February 2026 (mean@8) | 83.33 | 71.21 | 83.33 | 91.29 |
| Science | ||||
| GPQA-Diamond (pass@1) | 86.00 | 82.30 | 83.50 | 86.52 |
| HLE (128K context) | 21.40 | 20.60 | 21.83 | 24.79 |
| Logic | ||||
| BBEH (pass@1) | 48.05 | 55.18 | 64.69 | 52.51 |
| Code | ||||
| LiveCodeBench v6 (pass@1) | 85.32 | 85.96 | 76.80 | 85.66 |
It achieves the highest scores in competitive mathematics (AIME 2025, HMMT February 2026), GPQA-Diamond, and HLE. The margin is large against the second place in mathematics, while GPQA-Diamond is a close call. LiveCodeBench v6 is 85.66, falling short of Gemma-4’s 85.96. The difference is small, placing them on a similar level.
There are also weaknesses. IFEval is 93.00, failing to reach Muse Glimmer-30B’s 96.80. BBEH is 52.51, trailing behind Muse Glimmer-30B and Gemma-4. RULER also falls short of Qwen3.6’s 97.08.
Degradation Due to Quantization
In the GGUF version card, AtomicChat itself measures the difference in next-token distribution from BF16 (mean KLD) and the proportion of selecting the same token as BF16 (same top token).
| File | mean KLD | same top token |
|---|---|---|
Q8_0 |
0.0371 | 93.6% |
AD-Q6_K |
0.0482 | 92.6% |
AD-Q6_K-Q5_K |
0.0697 | 90.9% |
AD-Q5_K-Q4_K |
0.1383 | 87.3% |
AD-Q4_K-IQ4_XS |
0.1736 | 85.7% |
The card states that this model is sensitive to rounding and susceptible to a larger impact from quantization. While models in the same class typically maintain over 95%, it stays at 85.7% at 4.6 bits per weight. Configurations under 20GB were reportedly not published because the same top token dropped below 80%.
Even Q8_0 includes 0.037, which accounts for the lower bound caused by rounding error. The numbers above include this lower bound.
Strengths and Use Cases
Note that the contents of this article have not been officially confirmed, and the following is a report based on the descriptions in the original model Maxwell-V1-Flash card and the GGUF version card.
The original model card explains that it is designed to target reasoning, agentic, and long-context capabilities at practical inference costs. Post-training reportedly included an RL branch for agentic tasks such as search, terminal operation, function calling, and software engineering. The GGUF version tags also include agentic, and tool calling and reasoning are embedded in the chat template.
- Math and Science Reasoning: As shown in the table above, scores surpassing comparison targets are reported for math benchmarks like AIME 2025 and HMMT February 2026.
- Long Context: Supports up to 256K tokens. However, the GGUF version does not implement block-sparse attention as described below, behaving differently from the original beyond 16K.
- Multilingual: The tokenizer was reportedly built with consideration for low-resource Southeast Asian languages including Thai, Lao, Myanmar, Khmer, and Shan. The GGUF version card notes that Thai responses were manually verified using
llama-server --jinja. - Tool Calling: It is reported that reasoning and tool calling are interpreted via
llama-server --jinja.
There are also points to keep in mind. According to the GGUF version card, this model is prone to quantization impact. Furthermore, block-sparse attention and MTP (multi-token prediction) heads are unsupported in the GGUF version. For inputs exceeding 16K tokens, whereas the original model references only the blocks it selected, the GGUF version references the entire context.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 36.8B parameters (taken from the base model maxwell-model/Maxwell-V1-Flash)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 24GB (RTX 4090 / 3090, etc.) | IQ4_XS | 19.4GB | 23.2GB |
| 32GB (RTX 5090, etc.) | Q5_K | 25.3GB | 30.4GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 35.8GB | 42.9GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp as-is.
It is distributed in GGUF, so no conversion is needed.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compression: the IQ4_XS build measures 4.53 bits per weight — about 28% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| LiquidAI/LFM2.5-24B-A2B-Base | 24.0B | 80GB | lfm1.0 | LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-FP8 | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-bf16 | 24.0B | 80GB | lfm1.0 | LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-8bit | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B | 24.0B | 12GB | lfm1.0 | LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17) |
How to Get It
- Distribution Format: GGUF (
AtomicChat/Maxwell-V1-Flash-GGUF). Converted by AtomicChat from the original BF16 weights of the Maxwell team. BF16 is split into two parts, and the first part should be specified when loading. - License: Apache-2.0, originating from the model. No gating (agreement to terms of use) is set.
- Quantization Types: Features
Q8_0alongside Atomic Dynamic (AD) methods:AD-Q6_K,AD-Q6_K-Q5_K,AD-Q5_K-Q4_K, andAD-Q4_K-IQ4_XS. AD is described as a method that keeps non-routed experts at Q8_0 and quantizes only the experts with imatrix. - Supported Engines: Requires llama.cpp (commit
d3c3e242c) patched withmaxwell-moe-port.patchfrom the repository. It is noted that it cannot yet be loaded with standard llama.cpp, LM Studio, or Ollama. vLLM is mentioned in the original model card.
The build instructions follow the description in the card.
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git checkout d3c3e242c16b8cbf51400fad410cd69b22d64f35
git apply /path/to/maxwell-moe-port.patch
cmake -B build -DGGML_CUDA=ON
cmake --build build -j --config Release --target llama-server llama-cli
On a Mac, omit -DGGML_CUDA=ON (Metal is enabled by default). Here is an example startup command. Sampling settings are indicated by the original model publisher.
./build/bin/llama-server -m Maxwell-V1-Flash-AD-Q4_K-IQ4_XS.gguf --jinja -c 32768 -ngl 99 \
--temp 0.6 --top-p 0.95 --top-k 20
Thinking is enabled by default. To disable it, send "chat_template_kwargs": {"enable_thinking": false} in the request. Alternatively, specify --chat-template-kwargs '{"enable_thinking": false}' when launching the server.
Other Models for the Same Task
Recent text generation models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.
- plamo-3-8b-fin-instruct Text Generation Model: 4GB+ VRAM, GGUF Builds (8.5B, 4GB)
- Qwen3.8-27B-DSpine-b16 Draft Model for Speculative Decoding: 8GB+ VRAM (2.0B, 8GB)
- Arete-27B Text Generation Model: Our Test Answers, 12GB+ VRAM (27.4B, 12GB)
- ByteDance-Seed Releases Cola2 Language Model (8.8B, 48GB)
See all text generation models →
What to Read Next
- Find models by VRAM (This model runs from the 24GB tier) → Other models that run on a 24GB GPU
- Engines that run this model → llama.cpp / Ollama
- What IQ4_XS, Q5_K, Q8_0 mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- How to read IFEval, LongBench, RULER → Benchmark glossary
Sources
- https://huggingface.co/AtomicChat/Maxwell-V1-Flash-GGUF
- https://huggingface.co/maxwell-model/Maxwell-V1-Flash
- https://huggingface.co/datasets/AtomicChat/Maxwell-V1-Flash-GGUF-metrics
This article contains unverified information. We will append an update note once it is confirmed by a primary source.
