Mamba2-primed-HQwen3-8B-Instruct

Mamba2-primed-HQwen3-8B-Instruct is a Hybrid language model consisting of 50% Attention layers and 50% Mamba2 layers, primed from Qwen3-8B using the Hybrid Model Factory Priming pipeline. The model is instruction-tuned and supports context lengths up to 128K tokens.

Mamba-2 is a State-Space Model layer with constant memory and linear compute cost in the sequence length.

By combining Attention with Mamba-2, our Hybrid model achieves up to 2× faster inference at long contexts while closely matching the base Transformer's quality.

Links

• 📄 Mamba-2 paper
• 💻 GitHub: Hybrid Model Factory

Why Hybrid?

Each Primed Hybrid model is initialized from a base Transformer by converting a portion of its Attention layers into State-Space Model (SSM) layers that maintain a fixed-size recurrent state instead of a growing KV cache. At a 50% Hybrid ratio, roughly half the KV cache (which grows linearly with sequence length) is replaced with fixed-size SSM state. The practical benefits:

• Higher throughput at long contexts — less memory on KV cache means more memory for batching
• More concurrent sequences — ~2× as many concurrent sequences before hitting memory limits
• Growing advantage with context length — at long contexts, Attention dominates the forward pass while SSM layers remain negligible in cost. Since the Hybrid model makes roughly half as many Attention calls as the base Transformer, the throughput advantage grows with context length

Increasing hybridization ratio, replacing more Attention layers with SSM layers, further reduces memory and increases throughput, typically at the expense of performance.

Model Overview

• Type: Causal Language Model (Hybrid Attention + SSM)
• Base Model: Qwen3-8B
• Hybrid Layer Type: Mamba-2
• Hybrid Ratio: 50% (18 Attention + 18 Mamba-2 layers)
• Parameters: ~8B
• Context Length: 128K natively
• Precision: bfloat16
• License: Apache 2.0

Note, this is an Instruction-tuned model and is not a thinking model, that is, it does not natively produce chain-of-thought thinking tokens in its generation trace. Tool calling is not supported by this checkpoint. If you need tool calling capabilities, use the Reasoner checkpoints.

Benchmark Results

Below we report benchmark performance for all our instruction-tuned Primed models. All Hybrid models use a 50% Hybrid ratio and are Primed from Qwen3-8B.

We consider two baselines:

1. Qwen3-8B (non-thinking, from HF): The original Qwen model evaluated in non-thinking mode, which is the intended mode for an Instruct model. This serves as the base Transformer from which we start the Priming procedure.
2. Qwen3-8B (Long): The Qwen model fine-tuned on our priming data, extending its native context length from 32K to 128K. All Primed Hybrid models use the same training hyperparameters and data as this baseline, making it a fair comparison for differing architectures.

On both long- and short-context benchmarks, our Primed Hybrid models closely match the performance of the Transformer model while having considerably lower deployment costs, showcasing the efficacy of the Priming process.

Long-Context Benchmarks

Evaluated on HELMET, MRCR, and BABILong across context lengths from 8K to 128K, using a weighted average with geometrically increasing weights for longer contexts.

The plot below shows performance averaged over context lengths from 8K to 128K.

[!NOTE] For the Qwen3-8B (non-thinking, from HF) model, we used YaRN to evaluate on long-context tasks as directed in the model card

How close are the Hybrid models to the Transformer baseline on long context tasks? Primed GKA and GDN Hybrids are within ~1.5 points of Qwen3-8B (Long) on average, while being 1.5–2× faster at inference. Primed B'MOJO-F matches GKA/GDN in quality but is slower due to unfused SSM+SWA kernels (details). Primed Mamba2 lags further behind (approx. 3 point gap), consistent with GKA and GDN's higher expressivity.

Why SSM layers over Sliding Window Attention (SWA)? All Hybrid SSM models outperform the Hybrid SWA model (50% Attention + 50% SWA, window size 512). Even though SWA uses ~2× the effective state size of GKA at BF16, SSM layers retain information from the remote past, while SWA forgets everything beyond its window.

Short-Context NLP Benchmarks

Evaluations on Tulu3-dev from OLMES. All tasks are over a short-context length (≤ 8K). Each category in the table below averages the following Tulu3-dev subtasks:

1. Math: GSM8K, MATH.
2. Knowledge: MMLU, PopQA, TruthfulQA.
3. Coding: HumanEval, HumanEval+.
4. Reasoning: BigBenchHard.
5. Instruction Following: IFEval.

How close are the Hybrid models to the Transformer baseline on short context tasks? All Primed Hybrid models are within ~3 points of Qwen3-8B (Long), using <0.5% of the base Transformer's pre-training token budget. Note, B'MOJO-F fully matches the Transformer baseline but is slower to deploy (see above).

Which SSM layer type performs best? Among the non B'MOJO-F Hybrids, GKA ranks first (~2 point gap with Qwen3-8B [Long]), followed by GDN, then Mamba2. This ranking correlates with the expressiveness order of their respective SSM layers.

[!NOTE] For applications to complex reasoning and coding problems check out our Primed Hybrid Reasoning models.

About Mamba-2

Mamba-2 is a State-Space Model layer with diagonal transition dynamics and input-dependent gating. It processes sequences in linear time with constant memory, making it efficient for long-context inference. Mamba-2 is less expressive than Gated DeltaNet and Gated KalmaNet due to its diagonal structure, but benefits from well-optimized kernels.

For more details, see the Mamba-2 paper.

Architecture Details

Inference Efficiency

Sustained decode throughput (tokens/s) on 8× H200 GPUs (TP=8), measured during pure decode with a saturated KV cache. Benchmarked with random data (no prefix-caching benefits). See the full Inference guide for methodology and additional models.

Mean TTFT at the Transformer's saturated batch size (Hybrid model has memory to spare):

The decode throughput advantage grows with context length — from 1.88× at 16K to 2.30× at 128K — thanks to Mamba2 layers maintaining a fixed-size recurrent state instead of a growing KV cache. TTFT crosses over at 32K and reaches 0.74× (26% faster) at 128K.

Usage

With vLLM (recommended)

Install the Hybrid Model Factory vLLM plugin in your local environment, then serve:

vllm serve amazon/Mamba2-primed-HQwen3-8B-Instruct \
  --enable-prefix-caching \
  --mamba-cache-mode align \
  --mamba-cache-dtype float32 \
  --mamba-ssm-cache-dtype float32

Query the server:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amazon/Mamba2-primed-HQwen3-8B-Instruct",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "What is Linear Attention in the context of LLMs?"}
    ]
  }'

[!TIP] The --mamba-cache-dtype float32 and --mamba-ssm-cache-dtype float32 flags are important for accurate long-context generation. See the Inference guide for details on all recommended flags.

With Hugging Face Transformers

See the Inference guide for details on when we recommend the Hugging Face Transformers implementation as opposed to the highly optimized vLLM one.

from transformers import AutoModelForCausalLM, AutoTokenizer
import hmf.model.hybrid_zoo.models.model_register  # Register Hybrid models

model = AutoModelForCausalLM.from_pretrained(
    "amazon/Mamba2-primed-HQwen3-8B-Instruct", trust_remote_code=True
).to("cuda")
tokenizer = AutoTokenizer.from_pretrained("amazon/Mamba2-primed-HQwen3-8B-Instruct")

messages = [{"role": "user", "content": "What is linear attention in the context of LLMs?"}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training-Free Context Extension

This model supports training-free context extension 2-4× its native context via an extension to Hybrid models of PICASO cache composition. See the State Composition guide for usage. Note, this is currently supported in Hugging Face Transformers only.

Training data

These models were produced through the multi-stage Priming pipeline from Hybrid Model Factory. Training data spans web documents, mathematics, long-context documents, and instruction-following and reasoning examples — each targeting a different capability axis. This diversity is critical: it allows the Priming procedure to convert a base Transformer into a more memory- and compute-efficient Hybrid architecture at nearly the same level of performance, using <0.5% of the base Transformer model's pre-training token budget.

Responsible AI Considerations

At Amazon, we are committed to developing AI responsibly and take a people-centric approach that prioritizes education, science, and our customers, to integrate responsible AI across the end-to-end AI lifecycle. We believe the use of AI must respect the rule of law and human rights, and we encourage the safe and responsible development of AI. When downloaded or used in accordance with AWS Responsible AI Policy, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. Please report model quality, risk, security vulnerabilities or Amazon AI Concerns here.

Citation

@software{hybrid_model_factory,
  title = {Hybrid Model Factory},
  year = {2026},
  url = {https://github.com/awslabs/hybrid-model-factory}
}

@misc{dao2024transformersssmsgeneralizedmodels,
      title={Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality}, 
      author={Tri Dao and Albert Gu},
      year={2024},
      eprint={2405.21060},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2405.21060}, 
}

License

This model is licensed under the Apache 2.0 License.