Model card: sources and credits, veloGB10 TP=1/2/4 instructions, measured numbers, IP note, original card

77b05ee verified 4 days ago

21.6 kB

license: mit
license_link: https://huggingface.co/NaiveAI/Naive-N0.5-Flash/blob/main/LICENSE
model_name: doth4580/Naive-N0.5-Flash-EXL3-3.5bpw
base_model: NaiveAI/Naive-N0.5-Flash
base_model_relation: quantized
tags:
  - exl3
  - exllamav3
  - trellis
  - 3.5bpw
  - vllm
  - moe
  - long-context
  - dgx-spark
  - naive-n0.5-flash

Naive-N0.5-Flash — EXL3 3.5 bpw

Original model: NaiveAI/Naive-N0.5-Flash by the NaiveAI Team, MIT License. This artifact is quantized from the public FP8 release NaiveAI/Naive-N0.5-Flash-FP8. It contains modified weights (EXL3 trellis quantization); the original model is © its authors. Tokenizer and chat template are unchanged from the original. Served by the veloGB10 engine it is listed by /v1/models as doth4580/Naive-N0.5-Flash-EXL3-3.5bpw (the model_name: line of this card); base_model: is the original model it was quantized from, not this pack.

An EXL3 (ExLlamaV3 trellis) 3.5 bpw quantization of Naive-N0.5-Flash — 309B parameters, 15.5B active, native 1M-token context. The pack is ~~137 GB (~~128 GiB) against ~315 GB for the FP8 original, which is what lets it fit across two NVIDIA DGX Spark (GB10) machines.

Quantized with the exllamav3 tooling, adapted for this architecture, run on 8×H100 for speed. It is served on two Sparks with the recipe in sf-stav/Naive-N0.5-Flash-EXL3-2x-DGX-Spark — see Serving below.

What this pack is: sources and credits

Original model: NaiveAI/Naive-N0.5-Flash — © the NaiveAI Team, MIT; a copy ships here as LICENSE. Built on Xiaomi's MiMo-V2.5 and DeepSeek Sparse Attention.
This artifact: EXL3 3.5 bpw quantization produced with exllamav3 via the vcruz305 fork. Weights are modified; tokenizer and chat template are not.
Recipe: sf-stav/Naive-N0.5-Flash-EXL3-2x-DGX-Spark (MIT). Note that vllm-exl3 is AGPL-3.0 — relevant if you expose the server to others.

Running it with veloGB10

Download a release binary of veloGB10 (one self-contained file; see its README), and the pack with hf download doth4580/Naive-N0.5-Flash-EXL3-3.5bpw --local-dir /path/to/Naive-N0.5-Flash-EXL3-3.5bpw. The server is OpenAI-compatible on http://<head-ip>:9000/v1. The flags below are the ones this pack was tested with; --max-batch sets how many requests run at once.

TP=2 and TP=4 only: no single GB10 holds this model (the engine says so at boot).

Two GB10s (TP=2)

On the second machine (no model copy, no configuration):

./gb10_inference --node --port 29500

On the head (the machine with the model folder):

./gb10_inference --server \
  --model-dir /path/to/Naive-N0.5-Flash-EXL3-3.5bpw \
  --tp 2 --nodes <node-ip>:29500 \
  --host 0.0.0.0 --port 9000 \
  --max-seq-len 262144 \
  --max-batch 2 \
  --max-tokens 65536 \
  --prefix-cache on

Four GB10s (TP=4)

Start ./gb10_inference --node --port 29500 on each of the three other machines, then on the head:

./gb10_inference --server \
  --model-dir /path/to/Naive-N0.5-Flash-EXL3-3.5bpw \
  --tp 4 --nodes <node1-ip>:29500,<node2-ip>:29500,<node3-ip>:29500 \
  --host 0.0.0.0 --port 9000 \
  --max-seq-len 262144 \
  --max-batch 2 \
  --max-tokens 65536 \
  --prefix-cache on

The nodes need no copy of the model. The head sends each node only its own share of the weights over the network and keeps it in a local cache, so a second start of the same model transfers nothing. On a box with two active RoCE port pairs, name the one to use with --rdma-dev <rail1>,<rail2>; with one pair it is found automatically.

With the Naive S8 FP8 drafter (NaiveAI/Naive-N0.5-Flash-FP8-Draft): add --draft-dir /path/to/Naive-N0.5-Flash-FP8-Draft to the head's command line (any topology above).

Measured with veloGB10

VeloBench runs against the served API on NVIDIA DGX Spark-class GB10 machines (2026-11-09, veloGB10 0.7.7 development builds), at production sampling: each request sends the pack's default sampler settings, thinking on, one request at a time for the single-request columns. Decode is in output tokens per second and includes speculative decoding. One run per cell; sampled runs vary by roughly ±10%.

Topology	Draft source	Code turn decode	Story turn decode	All-turn median decode	Prefill (tok/s)	2 requests, aggregate	Draft acceptance
TP=2	Naive S8 FP8 drafter	40	31	33	761	46	28.9%
TP=4	Naive S8 FP8 drafter	80	46	50	1,077	69	34.7%

Accuracy

Measured 2026-10-19 with a forward-only gate in exllamav3 (teacher-forced NLL on a 977-token technical text; next-token probability after two code/JSON prefixes), the unquantized model through the authors' transformers code as the reference, same prompts for every row:

model	source	NLL (nats/token)	p(8-space indent after `if n < 0:`)	p(`age`) on a JSON key
Naive-N0.5-Flash bf16, `transformers`	bf16	1.627	1.000	1.000
this pack (3.5 bpw)	FP8	1.619	1.000	1.000
sibling EXL3 4 bpw (experts 4 / attention 6)	bf16	1.634	1.000	1.000

On this text the pack is indistinguishable from the unquantized model (the FP8 release it was made from scores marginally below the bf16 master). Agentic tool use, measured with tool-eval-bench v2.7.1 (69 scenarios, temperature 0, seed 42, thinking on) on this pack served by the veloGB10 engine at TP=4: 86 / 100 (118 / 138 points), with the layer-47 correction below applied. A runtime without that correction scores ~2.02 nats on the same text.

Runtime requirement (all EXL3 packs of this model): layer 47's routed up_proj is stored divided by 128 (exllamav3's interm_div for this architecture — that layer's expert activations overflow f16) and the forward must scale that layer's routing weights by 128, as exllamav3's BlockSparseMLP does. A runtime that applies the pack as stored produces a last-layer MoE 128× too small; the symptoms are repetition loops, lost indentation and garbled tool calls, not a crash. The veloGB10 engine had exactly this defect until 2026-10-19, and the vLLM recipe's vllm-exl3 path shows the same degraded NLL, so treat it as affected until its fix lands.

Specifications


Base model	NaiveAI/Naive-N0.5-Flash (FP8 release as the direct source)
Architecture	MoE, 309B total / 15.5B active, 48 layers, hidden 4096, 64 heads / 4 KV heads, head_dim 192
Attention	hybrid 39 SWA layers (window 128) + 9 DSA layers (top-2048 selection), GQA4, no full attention
MoE	256 routed experts, 8 per token, expert intermediate 2048
Context / vocab	1,048,576 tokens / 152,576
Quantization	EXL3 trellis, `mul1` codebook — uniform 3.5 bpw, 6-bit `lm_head`, embeddings unquantized, calibration 250×2048
Format / size	`exl3` (`--quantization exl3`) — ~~137 GB (~~128 GiB), 24 shards
License	MIT (inherited from the original model)

Files

File	Notes
`model-00001..00024-of-00024.safetensors`	the quantized weights
`model.safetensors.index.json`, `quantization_config.json`	tensor index and the full EXL3 metadata
`config.json`	carries the EXL3 `quantization_config` incl. the `non_routed_exl3` map (the vLLM-rewritten pack config)
`config.json.native`, `model.safetensors.index.json.native`	the equivalent native ExLlamaV3 pack metadata
`modeling_naive_n05_flash.py`, `configuration_naive_n05_flash.py`	model code, loaded via `trust_remote_code`
tokenizer + `chat_template.jinja`, `generation_config.json`	unchanged from the original

Notes

Beyond 2,048 tokens the DSA layers are approximate. The model's 9 DSA layers attend to the top-2,048 tokens picked by a learned indexer; the recipe runs them as dense attention, which is exact up to 2,048 tokens but past that sees more than the reference model does. Long-context answers may differ from the original; the 1M native context needs a sparse attention backend that is not implemented yet.
This is an early recipe, not a tuned one. Measured decode on 2× DGX Spark: ~~34.9 tok/s on pure code, ~23.9 mixed, ~19.5 reasoning (~~23.5 without speculation). Performance tables, the profile, and the limitations are in the recipe's README.
One request at a time, bf16 KV only, and keep the Sparks otherwise idle (GPU_MEM_UTIL=0.80).

Intellectual property

We do not claim any intellectual-property ownership of the model, its weights, its architecture or its original documentation. All rights stay with the original authors and with every contributor credited above, under their own licences, which apply to this repository unchanged. Our contribution is limited to what the "What this pack is" section says we did (quantization, packaging, testing and the veloGB10 instructions). If you are a rights holder and want something here changed or removed, open a discussion on this repository.

Original model card — NaiveAI/Naive-N0.5-Flash

Reproduced verbatim from the original repository (metadata header removed).

Building Frontier AI with AI

🏠 Homepage · 📰 Technical Blog · 💻 GitHub

Introduction

Naive-N0.5-Flash is an open-weight 309B MoE model with 15.5B active parameters, built for coding and AI R&D. It supports a native 1M-token context window through a hybrid of Sliding-Window Attention (SWA) and lightweight DeepSeek Sparse Attention (DSA), with no full-attention layers.

Key Features

Native 1M context, without full attention. Naive-N0.5-Flash combines Sliding-Window Attention (SWA) and lightweight DeepSeek Sparse Attention (DSA) with GQA4 at a predominantly 5:1 SWA–DSA layout. The entire network remains local or sparse, with no full-attention layers.
AI-optimized inference up to 2,000 tokens/s. NaiveRT, our inference system for Naive-N0.5-Flash, was built and optimized through AI-centered R&D. It combines mega-kernel fusion, Programmatic Dependent Launch (PDL), and speculative decoding, delivering 50 tokens/s per user in Standard mode and up to 2,000 tokens/s in Ultrafast mode. See the NaiveRT case study in the technical blog for the implementation and optimization process.
Open weights and API. Model weights and inference code are released under the MIT license. API access will also be provided, with pricing set at $0.10 / $0.40 / $0.01 per million tokens for input, output, and cache reads, respectively.

Model Architecture

Property	Specification
Architecture	Mixture-of-Experts (MoE)
Total parameters	309B
Active parameters	15.5B
Context length	Native 1M tokens
Transformer layers	48
Attention-layer composition	39 SWA layers + 9 DSA layers
Attention mechanism	Hybrid SWA–DSA
SWA window	128 tokens
DSA token selection	Top 2,048 tokens for backbone attention
DSA KV groups	4 (GQA4)
Indexer query heads	16

Hybrid SWA–DSA Attention

Naive-N0.5-Flash builds on the open-weight MiMo-V2.5 base model, which has a simple architecture with strong foundational capabilities in world knowledge and deep research. Most layers use Sliding-Window Attention (SWA), whose per-token decoding cost does not grow with context length, while a small number of global-attention layers preserve long-range information. At million-token context lengths, however, these global-attention layers account for much of the decoding overhead.

Naive-N0.5-Flash replaces the global-attention layers with DeepSeek Sparse Attention (DSA). A lightweight indexer scores the full history, while the backbone computes attention only over a selected subset of tokens. Although the indexer still scans the full history and the full KV cache is retained, sparse attention substantially reduces attention computation and memory access. Adapting the model to this new attention structure was one objective of continued pretraining.

Figure 1. The hybrid attention stack and DSA module.

The network consists of eight six-layer modules. A standard module contains five SWA layers followed by one DSA layer, with the first layer of the first module also replaced by DSA. SWA uses a 128-token window, while DSA selects the top 2,048 tokens for backbone attention. Both attention types incorporate sink bias.

Unlike the original MLA-based DSA implementation, Naive-N0.5-Flash replaces MLA with grouped-query attention (GQA) using four KV groups. For the architecture design process and indexer efficiency comparison, see model architecture in the technical blog.

Training Overview

Following the architectural changes, Naive-N0.5-Flash completed 3.25T tokens of multi-stage training with a native 1M-token context window: 50B tokens of Indexer Warmup, 3T tokens of Sparse Attention Training, and 200B tokens of Learning Rate Decay. This process adapted the model to its new sparse attention architecture while substantially improving its AI R&D and coding capabilities. See the technical blog for training details.

Evaluation Results

Figure 2. Coding and agentic task results. Naive-N0.5-Flash is highlighted in yellow.

Figure 3. AI research and systems optimization results. Metric directions are indicated in the figure.

Evaluation setup and metric notes

Evaluation setup. Unless otherwise noted, our evaluations of Naive-N0.5-Flash use Claude Code 2.1.207 with a 1M-token context window, temperature 1.0, and top-p 0.95. The harness exposes only basic file I/O and Bash tools.

Sources for reported benchmark scores are as follows:

GLM-5.3 and GLM-5.3-Flash: GLM-5.3 blog and GLM-5.3-Flash blog, respectively.
Kimi-K3: Kimi-K3 model page.
Qwen-3.8-Max: Qwen-3.8-Max blog.
Hy4-preview: Hy4-preview model page.
DeepSeek-V4.1-Flash: DeepSeek-V4.1-Flash model page.
Step-5-preview: Step-5-Preview-BF16 model page.
Fable-5 (w/ fallback): GLM-5.3 blog.
SWE-Bench Pro: GPT-5.6-Sol, Opus-5, and Opus-5.5 scores are drawn from the GPT-5.6 blog and the Claude Opus 5.5 System Card.
DeepSWE v1.1: The Muse-Spark-1.3 score comes from its Muse-Spark-1.3 blog. GPT-5.6-Sol and Opus-5 scores come from the DeepSWE v1.1 leaderboard. The Opus-5.5 score comes from the Claude Opus 5.5 System Card.
Terminal-Bench 2.1: Muse-Spark-1.3, GPT-5.6-Sol, and Opus-5 scores come from the Muse-Spark-1.3 blog. The GPT-6-Astra score comes from the Terminal-Bench 2.1 leaderboard.
ALE-CLI: GPT-5.6-Sol, GPT-6-Astra, Muse-Spark-1.3, Opus-5, and Opus-5.5 scores come from the ALE-CLI leaderboard.
FrontierSWE v1: We calculate the Dominance score using the competing systems’ results as of August 23, 2026.
ProgramBench: We report the Almost@1 score. GPT-5.6-Sol and Opus-5 scores come from the ProgramBench leaderboard.
MLE-bench-30: Gemini-3.5-Flash, Gemini-3.6-Flash, Grok-4.5, and GPT-5.6-Luna scores come from the Gemini 3.6 Flash model card. Following its evaluation protocol, we report the average position score of Naive-N0.5-Flash.
PaperBench: MiniMax M3, Opus-4.7, GPT-5.5, and Gemini-3.1-Pro scores come from the MiniMax M3 model page.
SOL-ExecBench, NanoChat AutoResearch, and NanoGPT SpeedRun: Naive-N0.5-Flash scores were obtained using our in-house AutoResearch harness. Recursive Superintelligence Inc. scores come from its research article. Following the SOL-ExecBench update, we use its updated score from the leaderboard.

Deployment

Naive-N0.5-Flash supports FP8 mixed-precision inference. For general use, we recommend setting the sampling parameters to temperature=1.0 and top_p=0.95.

Quick Start with Transformers

Naive-N0.5-Flash requires FP8-capable NVIDIA GPUs. The model weights occupy approximately 315 GB; allow additional GPU memory for inference.

pip install "transformers[torch,kernels]>=5.17.0"

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NaiveAI/Naive-N0.5-Flash-FP8"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype="auto",
    device_map="auto",
)

inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hello!"}],
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

output = model.generate(**inputs, max_new_tokens=2048)
response = tokenizer.decode(
    output[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(response)

License

Naive-N0.5-Flash is released under the MIT License.

Citation

If you find Naive-N0.5-Flash useful in your research or work, please cite:

@misc{naiveai2026naiven05flash,
  title  = {Naive-N0.5-Flash: Building Frontier AI with AI},
  author = {{NaiveAI Team}},
  year   = {2026},
  url    = {https://naive.ai/en/research/}
}

Acknowledgments

Naive-N0.5-Flash builds on the work of the open-source community and gives back to it. We thank the Xiaomi MiMo team for making their MiMo-V2.5 base model publicly available, the DeepSeek team for their work on DeepSeek Sparse Attention (DSA), and the SGLang team and community for their open-source inference infrastructure.

Contact

For questions, feedback, or collaboration, please contact us at contact@naive.ai or follow us on X at @naiveailab. You can also find our open-source projects and model releases on GitHub and Hugging Face.