- Status
- Verified
- Base model
- …
- Lineage
- …
- Unchanged from base
- …
- Trending
- #21
- Downloads, 30 days
- 1.7k
- Weights
- 73.6 GB
- Revision
Maxwell-V1.0-Flash is a next-generation, high-performance Mixture-of-Experts (MoE) language model developed by the Maxwell Team, with 35B total parameters and 3B active parameters per token (both excluding embeddings) and support for context lengths up to 256K tokens.
At a glance
- Task
- Text generation
- Input
- text
- Output
- text
- Parameters
- 36.8B, 256 experts, 12 active
- Architecture
- Maxwell MoE
- Context
- 512K tokens
- Precision
- BF16 100%
- Format
- Safetensors
- License
- Apache 2.0Commercial use
- Released
- Oct 2026
- Updated
- Nov 2026
- Likes
- 107
- Downloads, all time
- 287
Architecture
- Layers
- 45
- Hidden size
- 2,048
- Attention
- 32 heads, grouped query, 4 KV heads
- Experts
- 256 total, 12 active per token
- Vocabulary
- 200,704
- Positions
- 524,288
- RoPE theta
- 20,000,000
- Tied embeddings
- No
Family
Models built on Maxwell-V1-Flash.
Run it
Pinned to the indexed revision.
hf download maxwell-model/Maxwell-V1-Flash --revision ca536c4380e0600d6aebdb808d2ce65ec724161bThrough the hub: the same tools, each file from a source that is up (Hugging Face, ModelScope, IPFS), at this revision. The last line checks every file against the hash the tree names.
export HF_ENDPOINT=https://gethologram.ai
hf download maxwell-model/Maxwell-V1-Flash --quiet && hf cache verify maxwell-model/Maxwell-V1-FlashSpaces
Used in 1 Spaces.
Read the full model card

[](https://maxwell-llm.ai/)
[](https://maxwell-llm.ai/playground)
[](https://huggingface.co/maxwell-model/Maxwell-V1-Flash)
[](#6-license)
📄 [Technical Report](https://github.com/maxwell-model/Maxwell-V1-Flash/blob/main/Maxwell_v1_flash_tech_report.pdf) |
💬 [GitHub](https://github.com/maxwell-model/Maxwell-V1-Flash)
1. Model Introduction
Maxwell-V1.0-Flash is a next-generation, high-performance Mixture-of-Experts (MoE) language model developed by the Maxwell Team, with 35B total parameters and 3B active parameters per token (both excluding embeddings) and support for context lengths up to 256K tokens. It is designed for strong reasoning, agentic, and long-context capabilities at a practical inference cost.
Key Features
-
Maxwell Block Sparse Attention (MBSA). Each KV group keeps fixed coverage of 8 prefix and 8 local 128-token blocks and adds up to 16 remote blocks chosen by a learned indexer, bounding main attention at 4,096 tokens per KV group while the full-history KV cache stays available. The MBSA kernel is 6.7x faster than dense GQA at 256K.
-
Manifold-Constrained Hyper-Connections. Multiple residual streams with learned, input-dependent aggregation and constrained mixing, widening information flow without extra attention or feed-forward evaluations.
-
Fine-Grained Mixture of Experts. A mostly-MoE backbone pairing fine-grained routed experts with always-active shared experts, using a dynamic routing correction bias alongside a lightweight sequence-wise load-balancing loss.
Pre-training. Approximately 26T tokens of multilingual data, with the Muon optimizer on most hidden weight matrices. Training starts dense; block-sparse attention enters at 128K mid-training, once the indexer is warmed up against dense full-history attention, then trains jointly with the backbone.
Tokenizer and Long Context. A script-aware multilingual tokenizer stays compact across languages, especially low-resource Southeast Asian ones such as Thai, Lao, Myanmar, Khmer, and Shan. Long-context training then extends the sequence length in stages from 4K to 256K.
Post-training. Two-stage all-domain SFT builds general, reasoning, and agentic capabilities, then targets harder reasoning problems and agent trajectories. Independent RL branches cover general tasks, verifiable reasoning (mathematics, science, logic, code), and agentic tasks (search, terminal use, function calling, software engineering). Multi-teacher on-policy distillation then merges them into one model by supervising student-generated responses.
The results below summarize Maxwell-V1.0-Flash's performance across reasoning, agentic search, and other capabilities alongside the evaluated comparison models.

2. Model Specifications
Component
Configuration
Value
Model
Architecture
Mixture-of-Experts (MoE)
Model type
Text-only causal language model
Total / active parameters
35B / 3B
Tokenizer
Vocabulary size
200,704
Backbone
Transformer layers
45
Hidden size
2,048
Configured sequence length
262,144
Attention
Query heads / KV heads
32 / 4
Head dimension
128
RoPE base
10,000,000
Dense-GQA query-position prefix
16K tokens
Attention beyond the dense prefix
BSA
Block size
128 tokens
Maximum block budget per KV group
32
Prefix / local / remote block budgets
8 / 8 / 16
Indexer
Query heads / key heads
4 / 1
Head dimension
128
Dense FFN
Layers
3
Intermediate size
6,144
MoE
Layers
42
Routed experts (total / active per token)
256 / 12
Shared experts per layer
2
Intermediate size per expert
512
mHC
Residual streams
4
Sinkhorn–Knopp iterations
20
Initial scale of input-dependent mappings
0.01
MTP
Auxiliary layers
1
Attention type
Dense GQA
The block budget applies in sparse-attention mode. The MTP layer is auxiliary and is not included in the backbone layer count.
3. Evaluation Results
Base model. Maxwell-V1.0-Flash (Base) is compared with two Qwen base models of comparable scale under a unified internal evaluation framework. All scores are percentages; the best score in each row is shown in bold. GSM8K is evaluated using pass@8.
Category
Benchmark
Qwen3.0-w30a3-base
Qwen3.5-w35a3-base
Maxwell-V1.0-Flash (Base)
Math
GSM8K (pass@8)
97.95
98.41
97.19
MATH-500
59.00
57.50
71.40
MinervaMath (MATH)
62.14
58.62
74.56
MathQA
60.40
63.69
66.10
Code
HumanEval
70.73
67.68
78.05
MBPP
73.93
61.09
65.37
HumanEval+
63.41
56.10
71.95
MBPP+
75.40
70.90
75.13
STEM
MMLU
80.98
82.85
79.85
MMLU-Redux
78.17
78.17
74.90
MMLU-Pro
61.64
64.68
65.25
ARC-Easy
80.01
82.15
86.78
ARC-Challenge
57.59
62.12
65.78
BBH
28.08
52.17
57.01
GPQA-main
40.62
54.69
52.68
GPQA-diamond
42.42
54.04
51.52
Multilingual
MMLU (SEA)
73.89
77.37
75.06
XCOPA (SEA)
79.63
78.93
79.80
OpenBookQA (SEA)
43.20
41.30
43.30
ARC-Challenge (SEA)
51.83
54.67
58.13
MGSM
85.20
84.13
83.87
Knowledge
HellaSwag
81.38
83.64
79.84
PIQA
81.12
82.54
82.54
WinoGrande
77.74
74.11
72.30
TriviaQA
66.66
71.01
75.02
RACE
86.89
81.15
84.88
OpenBookQA
90.00
88.60
89.60
Chinese
CMMLU
87.41
86.17
79.41
C-Eval
88.78
86.33
77.41
Base model, long context. Scores are on a 0–100 scale. All results use block-sparse attention in every layer.
Benchmark
Context length
Qwen3.5-w35a3-base
Maxwell-V1.0-Flash (Base)
RULER
8K
91.98
94.80
16K
91.44
94.72
32K
90.91
94.63
64K
89.65
93.77
128K
88.39
94.57
256K
85.75
93.35
HELMET
8K
56.62
62.61
16K
54.93
61.43
32K
56.87
60.28
64K
54.61
58.74
128K
52.35
56.55
Post-trained model. Maxwell-V1.0-Flash is compared with four models of comparable scale: Qwen3.6-35B-A3B, which has the same nominal 35B total / 3B active parameter scale, together with Gemma-4-26B-A4B-it, Nemotron-3.5-Lightning, and Muse-Glimmer-30B. All scores are percentages; the highest score for each metric is shown in bold. Multilingual results cover only market languages.
Benchmark (Metric)
Qwen3.6
35B-A3B
Gemma-4
26B-A4B-it
Nemotron-3.5
Lightning
Muse
Glimmer-30B
Maxwell
V1.0-Flash
General
Instruction following
AdvancedIF (Score)
72.50
76.10
65.60
73.80
71.30
IFBench (Score)
59.60
79.10
70.80
83.20
64.80
IFEval (Score)
93.80
94.00
91.50
96.80
93.00
Instruction-following average (Average)
75.30
83.07
75.97
84.60
76.37
Multi-turn
MultiChallenge (pass@1)
62.59
65.76
61.07
70.45
72.20
MT-Bench-101 (Score)
94.20
93.90
94.50
88.20
93.30
MT-RAG (Average)
65.03
70.54
67.00
72.70
65.74
LongMemEval (Score)
79.52
85.89
83.44
94.48
86.18
Multi-turn average (Average)
75.34
79.02
76.50
81.46
79.36
Long context
AA-LCR (mean@5)
71.70
65.70
60.30
79.60
75.60
LongBench V2 (Score)
58.80
53.30
46.30
38.60
59.40
RULER (Score)
97.08
94.74
96.75
94.84
95.81
Long-context average (Average)
75.86
71.25
67.78
71.01
76.94
Multilingual
MLogiQA (Score)
77.8
77.8
71.3
80.0
81.5
CEval (Score)
88.3
83.7
76.2
85.4
91.0
MMMLU (Score)
89.1
90.9
87.8
90.7
92.0
INCLUDE (Score)
84.7
85.8
71.5
79.4
84.7
Belebele (Score)
93.8
94.8
91.8
94.3
92.8
Global-PIQA (Score)
91.7
93.9
90.9
92.4
92.3
MGSM (Score)
90.3
91.1
91.5
93.2
93.0
Multilingual average (Average)
88.0
88.3
83.0
87.9
89.6
Role playing
CoSER (pass@1)
47.79
56.19
43.78
60.32
53.54
RoleBench (pass@1)
89.21
84.05
81.10
73.40
89.92
Role-playing average (Average)
68.50
70.12
62.44
66.86
71.73
Reasoning
Mathematics
AIME 2025 (pass@8)
93.33
93.33
96.67
100.00
100.00
AIME 2025 (mean@8)
90.00
85.83
87.92
92.08
96.25
AIME 2026 (pass@8)
96.67
96.67
96.67
96.67
96.67
AIME 2026 (mean@8)
91.25
87.50
87.08
94.17
95.00
HMMT February 2025 (pass@8)
100.00
100.00
96.67
96.67
100.00
HMMT February 2025 (mean@8)
89.58
82.50
85.83
91.25
96.25
HMMT November 2025 (pass@8)
93.33
96.67
93.33
93.33
100.00
HMMT November 2025 (mean@8)
88.75
80.83
86.67
87.92
90.83
HMMT February 2026 (pass@8)
93.94
87.88
90.91
93.94
100.00
HMMT February 2026 (mean@8)
83.33
71.21
75.76
83.33
91.29
IMO-AnswerBench (pass@8)
92.25
88.50
88.00
91.75
95.50
IMO-AnswerBench (mean@8)
81.25
70.00
73.25
78.62
86.03
Mathematical average (pass@8)
94.92
93.84
93.71
95.39
98.70
Mathematical average (mean@8)
87.36
79.65
82.75
87.90
92.61
Science
GPQA-Diamond (pass@1)
86.00
82.30
75.44
83.50
86.52
MMLU-Pro (Accuracy)
85.20
82.60
81.94
83.16
85.46
MMLU-Redux (Accuracy)
93.30
93.10
91.91
90.72
93.59
Scientific average (Average)
88.17
86.00
83.10
85.79
88.52
HLE (128K context) (Score)
21.40
20.60
11.40
21.83
24.79
Logic
BBH (pass@1)
90.87
86.50
84.03
89.33
91.02
BBEH (pass@1)
48.05
55.18
25.86
64.69
52.51
KOR-Bench (pass@1)
74.48
77.36
72.32
80.08
77.20
Logical macro average (Macro average)
71.13
73.01
60.74
78.03
73.58
Code
LiveCodeBench v6 (pass@1)
85.32
85.96
78.30
76.80
85.66
LiveCodeBench v6 (pass@8)
92.32
92.89
87.90
85.70
93.74
Agent
Terminal
Terminal-Bench (mean@5)
47.42
38.43
23.10
48.50
58.65
SWE
SWE-Bench-Verified (Claude Code) (pass@1)
63.40
39.20
42.60
47.60
65.40
Tool use
MCP-Mark (mean@8)
36.70
14.50
14.47
27.56
37.30
Tau3-Bench (pass@1)
64.00
58.60
58.63
63.20
67.37
Search
BrowseComp (pass@1)
47.00
29.20
24.20
55.60
76.30 / 82.20*
HLE w/Tools (pass@1)
28.30
31.70
16.60
34.40
46.10 / 55.10*
WideSearch (max@4)
70.40
54.50
60.70
70.70
74.40 / 75.80*
DeepSearchQA (pass@1)
71.10
63.00
62.80
81.80
83.30 / 89.50*
Bamboogle (mean@8)
90.90
91.30
90.60
86.00
92.10
WebWalkerQA-300 (mean@8)
62.80
54.50
55.40
56.00
70.50
2WikiMultihopQA-300 (mean@8)
87.75
89.20
86.20
89.40
89.80
MuSiQue-300 (mean@8)
49.25
49.30
42.40
46.00
63.00
HotpotQA-300 (mean@8)
80.80
79.50
73.20
80.50
84.70
PopQA-300 (mean@8)
68.29
64.70
58.90
68.80
70.10
* Maxwell with context management enabled.
Post-training evaluations mostly use temperature 0.6, top-p 0.95, top-k 20, and a 32K-token context. AA-LCR, mathematical benchmarks, and HLE use 128K; scientific benchmarks use 80K; LiveCodeBench v6 uses 64K; multilingual evaluation uses a maximum generation length of 16K tokens. SWE-Bench Verified is evaluated with the Claude Code harness (version 2.1.212) in a network-isolated environment, with a 256K-token context window, at most 240 interaction turns, and a four-hour wall-clock limit per task. Terminal-Bench 2.1 is evaluated in the Harbor framework with the Terminus-2 harness, using a two-hour timeout per task, a 256K-token context window, 32K max tokens, temperature 1.0, top-p 0.95, top-k 20, and interleaved thinking.
The scientific average excludes HLE.
4. Deployment
Requires NVIDIA driver >= 580 (CUDA 13.0). On older drivers, add
-e VLLM_ENABLE_CUDA_COMPATIBILITY=1 to docker run.
Download the weights to a local directory first with the Hugging Face CLI
(install it with pip install -U huggingface_hub if hf is not available):
export MODEL_DIR=$PWD/Maxwell-V1-Flash
hf download maxwell-model/Maxwell-V1-Flash --local-dir "$MODEL_DIR"
Start the vLLM OpenAI-compatible server:
docker run --rm --init \
--name vllm-maxwell \
--gpus all \
--shm-size 8g \
--network host \
-v "$MODEL_DIR":/models/Maxwell-V1-Flash:ro \
--entrypoint vllm \
maxwellmodel/vllm-openai:v0.30.0 \
serve /models/Maxwell-V1-Flash \
--trust-remote-code \
--block-size 128 \
--served-model-name Maxwell-V1.0-Flash \
--port 8000 \
--tensor-parallel-size 1 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder
You can also try Maxwell-V1.0-Flash at https://maxwell-llm.ai/playground.
5. Model Usage
Maxwell-V1.0-Flash uses the bundled chat template for system, user, assistant,
and tool messages, with thinking enabled by default. For multi-turn
conversations and tool calls, pass the complete assistant message back into
messages, including reasoning_content and tool_calls when present.
OpenAI-compatible tool definitions are accepted through the tools field.
import openai
def chat_completion(
client: openai.OpenAI,
model_name: str,
) -> str:
messages = [
{
"role": "user",
"content": (
"Why is it harder to keep a room cool when the door is open "
"on a hot day? Explain in 3-4 simple sentences."
),
},
]
response = client.chat.completions.create(
model=model_name,
messages=messages,
stream=False,
max_tokens=4096,
temperature=0.6,
top_p=0.95,
extra_body={
"top_k": 20,
"chat_template_kwargs": {
"enable_thinking": True,
},
},
)
return response.choices[0].message.content
For full guides and examples, see the Maxwell-V1.0-Flash Quickstart.
6. License
Maxwell-V1.0-Flash is released under the Apache License 2.0. See LICENSE for details.
7. Contact Us
If you have any questions or suggestions, feel free to reach out to our R&D and product teams via email: contact@maxwell-llm.ai
Maxwell-V1.0-Flash is developed by the Maxwell team.
Derived on Nov 17, 2026 from Hugging Face at revision ca536c43, README.md , config.json .
50 files, 73.6 GB. Every download is checked against its address.