DEV.co
Open-Source LLM · Ex0bit

Qwen3.6-35B-A3B-PRISM-NVFP4

Qwen3.6-35B-A3B-PRISM-NVFP4 is a quantized 35-billion-parameter mixture-of-experts language model from Ex0bit, optimized for NVIDIA Blackwell GPUs. It combines PRISM tuning (which reduces refusal behavior and removes bias patterns) with aggressive FP4 quantization to fit in ~24 GB. It includes multimodal capabilities and is designed for vLLM inference.

Source: HuggingFace — huggingface.co/Ex0bit/Qwen3.6-35B-A3B-PRISM-NVFP4
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
39.4k
Downloads (30d)

Key facts

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

FieldValue
DeveloperEx0bit
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads39.4k
Likes20
Last updated2026-05-03
SourceEx0bit/Qwen3.6-35B-A3B-PRISM-NVFP4

What Qwen3.6-35B-A3B-PRISM-NVFP4 is

A W4A4 (4-bit weight, 4-bit activation) compressed-tensors quantization of Qwen/Qwen3.6-35B-A3B, a 256-expert MoE model with ~3B active parameters per token. Keeps vision encoder, lm_head, router gates, embeddings, and SSM state in BF16. Targets vLLM on SM120/SM121 (Blackwell) with specialized PTX kernels. PRISM tuning softens over-refusal and removes propaganda/bias. Includes optional MTP draft head for speculative decoding.

Quickstart

Run Qwen3.6-35B-A3B-PRISM-NVFP4 locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Ex0bit/Qwen3.6-35B-A3B-PRISM-NVFP4")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

Latency-sensitive, cost-constrained LLM serving

The W4A4 quantization and sparse MoE activation reduce memory footprint and compute per token, enabling affordable serving on mid-range Blackwell hardware without significant quality loss.

Bias and refusal reduction for content generation

PRISM tuning explicitly targets over-refusal behavior and propaganda/bias removal; suitable for applications requiring more permissive output (creative writing, open-ended dialogue) while maintaining coherence.

Multimodal AI assistants and RAG systems

Preserved vision encoder and multimodal capability make this suitable for image-to-text tasks, visual QA, and document understanding in RAG pipelines.

Running & fine-tuning it

Estimated ~24 GB peak memory (model card stated). Designed for NVIDIA Blackwell (H200, H100 variant, or DGX Spark with SM120/SM121). Requires custom vLLM kernels; stock PyPI builds insufficient. Community Docker images (aeon-7, avarok) available but require verification. Non-Blackwell serving falls back to slower software-E2M1 conversion or Marlin-MoE.

Unknown. Model card does not discuss fine-tuning, LoRA, or QLoRA feasibility. Quantized weights in NVFP4 format may complicate standard LoRA adapters. Requires experimentation or custom tooling. Not recommended for rapid iteration without prior quantized-model LoRA validation.

When to avoid it — and what to weigh

  • Strict safety/refusal guardrails required — PRISM explicitly softens refusal behavior and removes some safety constraints. If your application demands strong content filtering, this model's design philosophy conflicts with that requirement.
  • Non-Blackwell or non-NVIDIA hardware — Requires SM120/SM121 Blackwell tensor cores and custom vLLM kernels (the stock PyPI wheels will fault). Alternatives (Marlin-MoE, software E2M1) exist but are slower. Not compatible with non-NVIDIA accelerators or older NVIDIA GPUs without recompilation.
  • Deployment on older vLLM versions — Requires vLLM with compressed-tensors quantization support and Blackwell-specific patches. Older deployments or unpatched builds will not work.
  • Production quality assurance without domain validation — No published benchmarks, ablations, or safety assessments. Quantization and PRISM tuning impact are unknown without testing on your specific tasks.

License & commercial use

Apache 2.0, inherited from the base model Qwen/Qwen3.6-35B-A3B. Apache 2.0 is an OSI-approved permissive license with clear commercial use provisions.

Apache 2.0 permits commercial use, modification, and distribution under the terms of the license (attribute source, include license and CHANGES). However, confirm that the base model Qwen/Qwen3.6-35B-A3B is also Apache 2.0 and that no additional restrictions apply upstream. No licensing restrictions stated for PRISM tuning or quantization. Commercial usage for this specific quantization variant should be permissible, but verify base model licensing independently.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceModerate
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceMedium
Security considerations

PRISM tuning removes propaganda/bias patterns, but this is a design choice, not a security guarantee. FP4 quantization may introduce numeric instability in edge cases (unknown impact). Quantized models are less interpretable; adversarial robustness is untested. Custom vLLM kernels (PTX instructions) should be reviewed before deployment in sensitive environments. No vulnerability disclosures or security audits mentioned.

Alternatives to consider

Qwen/Qwen3.6-35B-A3B (unquantized)

Original base model; higher quality and no hardware lock-in, but ~140 GB+ footprint and higher inference cost. Better for environments with abundant VRAM and no strict latency constraints.

Llama-3.1-70B-Instruct (with GPTQ/GGUF quantization)

Larger, well-established baseline with broader hardware support (llama.cpp, TGI, GGUF ecosystem). Quantized variants available for cost reduction, but no MoE sparsity or PRISM tuning.

Mistral 7B or Mixtral 8x7B (with standard quantization)

Smaller, widely supported, easier deployment. Trade-off: fewer parameters and reduced multimodal capability, but proven in production and no custom kernel requirements.

Software development agency

Ship Qwen3.6-35B-A3B-PRISM-NVFP4 with senior software developers

Verify Blackwell GPU availability and vLLM kernel compatibility before production. Use provided Docker images or build custom kernels. Test PRISM tuning behavior on your workload. Contact our team for guidance on private deployment, fine-tuning, or model selection.

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.

Qwen3.6-35B-A3B-PRISM-NVFP4 FAQ

Can I use this model commercially?
Apache 2.0 license permits commercial use, but confirm the base model (Qwen/Qwen3.6-35B-A3B) is also Apache 2.0 and free of additional restrictions. Once confirmed, you may use, modify, and distribute under Apache 2.0 terms (attribute source, include license). Consult legal if commercial use is mission-critical.
What hardware do I need?
Minimum: NVIDIA Blackwell GPU (H200, H100 variant, or DGX Spark with SM120/SM121) and ~24 GB VRAM. Requires custom vLLM build with Blackwell NVFP4 kernels; stock PyPI wheels will fail. Use community Docker images (aeon-7, avarok) for faster setup, but verify compatibility with your cluster.
How does the PRISM tuning affect output quality?
PRISM softens over-refusal and removes propaganda/bias, but no published benchmarks or quality metrics are available. Expect more permissive outputs and different behavior than the base model. Test on your use case before production.
Can I run this on older GPUs or AMD hardware?
No. NVFP4 quantization and Blackwell tensor cores are required. Fallback (software E2M1 conversion) is significantly slower. AMD hardware is unsupported. Consider standard quantization (GPTQ, AWQ) or unquantized variants for broader hardware compatibility.

Software development & web development with DEV.co

Adopting Qwen3.6-35B-A3B-PRISM-NVFP4 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy This Model?

Verify Blackwell GPU availability and vLLM kernel compatibility before production. Use provided Docker images or build custom kernels. Test PRISM tuning behavior on your workload. Contact our team for guidance on private deployment, fine-tuning, or model selection.