GKA-primed-HQwen3-32B-Reasoner is a Hybrid language model consisting of 50% Attention layers and 50% Gated KalmaNet (GKA) layers, primed from Qwen3-32B using the Hybrid Model Factory Priming pipeline. The model is trained for long-context reasoning and supports context lengths of 128K tokens.
GKA (pronounced as gee-ka) is a State-Space Model layer inspired by the Kalman Filter that solves an online ridge regression problem at test time, with constant memory and linear compute cost in the sequence length.
By combining Attention with GKA, our Hybrid model achieves up to 2× faster inference at long contexts while closely matching the base Transformer's quality.
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.
We consider the following Transformer as a baseline:
Qwen3-32B (thinking, from HF): The original Qwen model evaluated in thinking mode, which is the intended mode for reasoning tasks. This serves as the base Transformer from which we start the Priming procedure.
Evaluations on math reasoning (AIME24/25), science (GPQA), coding (LiveCodeBenchv5, Scicode), tool-calling (BFCLv3/v4), and instruction-following (IFBench). Evaluations are done using the Nemo Evaluator SDK. We have provided the evaluation configuration examples/evaluation/nemo_reasoning_evals.yaml for reproducibility. Evaluations are done at 64K generation length.
Model
AIME24
AIME25
GPQA
LiveCodeBench-v5
BFCLv4 (minus web-search)
BFCLv3
IFBench
SciCode
Average
Qwen3-32B (thinking, from HF)
86.33
70.00
65.40
64.44
69.30
69.57
32.61
15.94
59.20
GKA-primed-HQwen3-32B-Reasoner
87.67
81.67
67.30
70.24
70.14
66.34
48.22
12.34
62.99
For BFCLv4, we remove the web-search subtask and weight each task by the number of entries (test examples) for that task.
How close is the Hybrid model to the Transformer baseline on complex reasoning tasks? Our Primed GKA Hybrid outperforms the Qwen3-32B (thinking, from HF) baseline by ~3.8 points on average, despite using <0.5% of the base Transformer's pre-training token budget. This is enabled by our multi-stage reasoning training pipeline coupled with higher throughput of Hybrid architectures.
Gated KalmaNet is a State-Space Model layer that is more expressive than both Mamba2 and Gated DeltaNet. GKA achieves this by employing the Kalman Filter to compute the optimal state at each time-step based on the entire past. In contrast, SSMs like Mamba2 and GDN rely on instantaneous objectives (that rely solely on the current input and loss estimate of the past) to compute their state.
Unlike other SSM-based hybrid layers, GKA gives you a runtime knob for trading compute against speed — with no retraining nor architecture changes. The num_iter parameter controls how many iterations the GKA solver runs during inference. No other hybrid layer type offers this: GDN and Mamba2 have fixed compute per layer, so their speed is fixed a priori. GKA lets you slide along the compute–latency curve per deployment, making it uniquely suited for scenarios where different endpoints or traffic tiers have different latency budgets.
For details on controlling GKA's compute–speed tradeoff at serving time via num_iter, see GKA Compute Control, and for more details on the modeling choices see the GKA paper.
As discussed above, GKA offers the unique ability to adjust the inference FLOPs by tuning the num_iter parameter. Here we summarize reasoning performance across different num_iter settings.
For most practical scenarios, we recommend setting num_iter=10 for the best trade-off. See next section for inference gains upon reducing number of iterations.
Note. Interestingly, setting num_iter=0 effectively converts the GKA model to a Gated Linear Attention (GLA) model. Thus, one can think of increasing num iters as improving upon the initial solution of the GLA model.
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.
The decode throughput advantage grows with context length — from 1.29× at 16K to 1.99× at 128K (2.06× with num_iter=1) — thanks to GKA layers maintaining a fixed-size recurrent state instead of a growing KV cache. TTFT crosses over at long contexts: GKA prefills 10–16% faster than the Transformer at 128K depending on num_iter. Reducing num_iter progressively improves both decode and TTFT, with the effect more pronounced at 32B than 8B. See Trade-off inference FLOPs for accuracy for details.
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "amazon/GKA-primed-HQwen3-32B-Reasoner",
"messages": [
{"role": "user", "content": "What is Linear Attention in the context of LLMs?"}
],
"temperature": 1.0,
"top_p": 1.0
}'
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.
Tip. Similarly to NVIDIA-Nemotron-3-Nano-30B-A3B-BF16, for generic reasoning tasks (e.g. Math, Science) we recommend setting temperature=1.0 and top_p=1.0. For tool-calling we recommend temperature=0.6, top_p=0.95.
Our reasoning model supports thinking on/off modes. Whenever thinking mode is on, the model will reason for multiple tokens in a segment delimited by <think> and </think> (which is extracted by the reasoning parser) before producing a response. This is necessary for difficult queries and increases response quality at the expense of higher latency. Thinking mode is enabled by default, however thinking mode can be turned off via the chat template.
If you want to query the model with thinking mode off, query the model as follows:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "amazon/GKA-primed-HQwen3-32B-Reasoner",
"messages": [
{"role": "user", "content": "What is Linear Attention in the context of LLMs?"}
],
"chat_template_kwargs": {"enable_thinking": false}
}'
Warning. Due to the long generations produced by reasoning models, the lower latency provided by vLLM is preferred over Hugging Face for evaluations and in production settings. We recommend Hugging Face generation primarily for quick debugging or testing.
from transformers import AutoModelForCausalLM, AutoTokenizer
import hmf.model.hybrid_zoo.models.model_register # Register Hybrid models
model = AutoModelForCausalLM.from_pretrained(
"amazon/GKA-primed-HQwen3-32B-Reasoner", trust_remote_code=True
).to("cuda")
tokenizer = AutoTokenizer.from_pretrained("amazon/GKA-primed-HQwen3-32B-Reasoner")
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, enable_thinking=True
)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=65536, temperature=1.0, top_p=1.0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
In order to turn thinking mode off, simply specify enable_thinking=False when applying the chat template:
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, enable_thinking=False
)
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.
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.
import os
os.environ["HF_ENDPOINT"] = "https://freeweights.allternit.com"
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("amazon/GKA-primed-HQwen3-32B-Reasoner", revision="ef4e0ee205468cbdc0512655fb0d8952268b4891")
model = AutoModelForCausalLM.from_pretrained("amazon/GKA-primed-HQwen3-32B-Reasoner", revision="ef4e0ee205468cbdc0512655fb0d8952268b4891", device_map="auto")
BAAI