DEV.co
Open-Source LLM · amazon

BMOJOF-primed-HQwen3-8B-Instruct

BMOJOF-primed-HQwen3-8B-Instruct is an 8B parameter instruction-tuned language model that replaces half its standard attention layers with B'MOJO-F hybrid layers (combining sliding-window attention with state-space models). It achieves up to 1.75× faster inference on long contexts (up to 128K tokens) while maintaining performance within ~1–3 points of the base Qwen3-8B model on standard benchmarks. Licensed under Apache 2.0, ungated, and developed by Amazon.

Source: HuggingFace — huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct
9.3B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
69.4k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developeramazon
Parameters9.3B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads69.4k
Likes2
Last updated2026-04-03
Sourceamazon/BMOJOF-primed-HQwen3-8B-Instruct

What BMOJOF-primed-HQwen3-8B-Instruct is

Hybrid Attention + SSM architecture with 36 layers (18 standard Attention + 18 B'MOJO-F). Each B'MOJO-F layer couples sliding-window attention (2048-token window) with a GKA-based state-space model within a single layer, replacing the growing KV cache with fixed-size recurrent state. Precision: bfloat16. Native context: 128K tokens. Derived from Qwen3-8B via Amazon's Hybrid Model Factory priming pipeline. Achieves ~1.5–2× inference speedup at long contexts versus base transformer, with ~2× concurrent sequence capacity before memory saturation.

Quickstart

Run BMOJOF-primed-HQwen3-8B-Instruct locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="amazon/BMOJOF-primed-HQwen3-8B-Instruct")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Long-document processing and retrieval

Handles documents up to 128K tokens with 1.5–2× faster inference than pure attention models, making it cost-effective for contract review, legal discovery, technical documentation analysis.

High-throughput inference deployments

Hybrid architecture reduces KV cache memory by ~50%, enabling ~2× concurrent sequences on the same hardware. Ideal for multi-user SaaS, API services, or batch processing at scale.

Private/on-premises LLM deployment

Apache 2.0 licensed, ungated, suitable for self-hosted inference on consumer/enterprise GPUs. Competitive quality with Qwen3-8B (Long) at lower deployment cost.

Running & fine-tuning it

ESTIMATE: ~16–24 GB VRAM for inference in bfloat16 precision on single GPU (e.g., A100 40GB, L40S 48GB). Batching and long-context (128K) workloads may require larger allocators. ~2× lower KV cache memory than base Qwen3-8B at equivalent batch size due to hybrid layers. Exact requirements depend on serving framework (vLLM, TGI) and batch size.

Model is instruction-tuned; further fine-tuning is feasible but not documented in card. LoRA/QLoRA adaptation likely compatible with HuggingFace ecosystem (safetensors format), though no explicit fine-tuning examples or guidance provided. Hybrid layer structure (half B'MOJO-F) may require custom LoRA adapter logic; verify framework support before production fine-tuning.

When to avoid it — and what to weigh

  • Chain-of-thought reasoning critical — Model card explicitly states this is NOT a thinking model; it does not natively produce chain-of-thought tokens. Avoid if internal reasoning transparency is required.
  • Exact quality parity required — Hybrid models show ~1–3 point gaps versus base Qwen3-8B (Long) on short-context benchmarks. If marginal performance loss is unacceptable, use the base transformer instead.
  • Cutting-edge SOTA reasoning/math tasks — Benchmarks show gaps in math (65.69 vs 64.56 baseline) and reasoning (76.42 vs 76.27). Specialized reasoning models may outperform on frontier tasks.
  • Production use without kernel optimization — B'MOJO-F layers use unfused SSM+sliding-window-attention kernels, slower than optimized alternatives (GKA, GDN). Inference speed advantage requires low-latency serving infrastructure.

License & commercial use

Apache License 2.0 (OSI-approved permissive license). Allows commercial use, modification, and distribution with proper attribution and liability disclaimer.

Apache 2.0 is a clear permissive OSI license. Commercial use, including incorporation into proprietary applications and SaaS offerings, is permitted provided the license and copyright notice are retained and any claims are disclaimed. No commercial use restrictions apply.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceModerate
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM security practices apply: prompt injection, jailbreaking, and adversarial input handling are not addressed in card. No stated vulnerability disclosure policy or security audit mentioned. Recommend standard ML security review (input validation, output filtering, rate limiting) before production deployment. Hybrid SSM component is less widely battle-tested than pure transformers; risk profile unknown.

Alternatives to consider

Qwen3-8B (base, non-hybrid)

Drop-in replacement if performance parity is critical and inference speed advantage is not essential. Simpler architecture, mature tooling, no kernel fusion dependency.

GKA-primed-HQwen3-8B-Instruct or Mamba2-primed-HQwen3-8B-Instruct

Alternative hybrid architectures from same Amazon Hybrid Model Factory. GKA ranks higher on benchmarks (~2 pt gap vs 1 pt); Mamba2 is simpler but slower. Choose based on throughput vs. quality tradeoff.

Llama 3.1 8B or Mistral 8B

Mature, widely deployed 8B alternatives. Trade long-context support (128K) for broader ecosystem maturity and community support; no performance advantage on short-context tasks.

Software development agency

Ship BMOJOF-primed-HQwen3-8B-Instruct with senior software developers

Ideal for long-document processing, high-throughput inference, and private LLM deployments. Compare performance and cost on your workload. Review GitHub (awslabs/hybrid-model-factory) and benchmark details. Contact your DevCo architect to test on your infrastructure.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

BMOJOF-primed-HQwen3-8B-Instruct FAQ

Can I use this model commercially?
Yes. Apache 2.0 is a permissive OSI-approved license allowing unrestricted commercial use. Retain the license header and copyright notice in any distribution.
What GPU do I need for inference?
Estimated 16–24 GB VRAM in bfloat16. A100 (40GB), L40S (48GB), or RTX 6000 Ada (48GB) recommended for batching. Hybrid design reduces KV cache by ~50% versus base Qwen3-8B, enabling ~2× concurrent sequences. Exact requirements depend on batch size, context length, and serving framework.
How does this compare to the base Qwen3-8B?
Faster inference (1.5–2× on long contexts), same quality on long-context benchmarks. On short-context tasks, 1–3 point performance gap (average ~71.27 vs 71.21). Lower deployment cost due to smaller KV cache and higher throughput.
Can I fine-tune this model?
Likely yes via LoRA/QLoRA with standard HuggingFace tools (safetensors format supported), but no official fine-tuning guide provided. Hybrid B'MOJO-F layers may require custom adapter logic; verify framework compatibility before production use.

Software development & web development with DEV.co

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If BMOJOF-primed-HQwen3-8B-Instruct is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Evaluate BMOJOF-primed-HQwen3-8B-Instruct for Your Use Case

Ideal for long-document processing, high-throughput inference, and private LLM deployments. Compare performance and cost on your workload. Review GitHub (awslabs/hybrid-model-factory) and benchmark details. Contact your DevCo architect to test on your infrastructure.