ADu2021/deepseek-v3.2-frontier
skills/skillxiv-v0.0.2-claude-opus-4.6/deepseek-v3.2-frontier/SKILL.md
Combines DeepSeek Sparse Attention (DSA) achieving O(Lk) complexity, scalable RL framework allocating 10%+ compute to post-training, and large-scale agentic task synthesis with 1,800+ environments. Matches proprietary models in reasoning; DeepSeek-V3.2-Speciale achieves olympiad gold.
$ install --global
skillsauthnpx skillsauth add ADu2021/skillXiv deepseek-v3.2-frontierInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 17, 2026, 5:33 AM14.8s1 file scanned
SKILL.md
- name:
- deepseek-v3.2-frontier
- title:
- DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
- version:
- 0.0.2
- engine:
- skillxiv-v0.0.2-claude-opus-4.6
- license:
- MIT
- url:
- https://arxiv.org/abs/2512.02556
- keywords:
- [large-language-models, sparse-attention, reinforcement-learning, tool-use, reasoning]
- description:
- Combines DeepSeek Sparse Attention (DSA) achieving O(Lk) complexity, scalable RL framework allocating 10%+ compute to post-training, and large-scale agentic task synthesis with 1,800+ environments. Matches proprietary models in reasoning; DeepSeek-V3.2-Speciale achieves olympiad gold.
Summary
DeepSeek-V3.2 presents three major technical breakthroughs: DeepSeek Sparse Attention (DSA) reducing complexity from O(L²) to O(Lk), a scalable RL framework allocating over 10% of pre-training compute to post-training, and a large-scale agentic task synthesis pipeline generating 85,000+ synthetic prompts across 1,800 environments. Together these innovations enable competitive reasoning performance on olympiad-level problems.
Core Technique
DeepSeek Sparse Attention (DSA): Uses a "lightning indexer" to select only top-k most relevant tokens per query. This reduces full-attention complexity dramatically while maintaining long-context quality.
Scalable RL Framework: Implements a robust post-training protocol using refined GRPO with: unbiased KL estimation preventing divergence and off-policy sequence masking for stability when training on diverse data.
Agentic Task Synthesis: Generates realistic multi-turn tool-use scenarios by creating 1,800 domains and 85,000 prompts requiring reasoning across tool interactions. Cold-start initialization seeds diverse reasoning patterns.
Implementation
DSA mechanism: For each query position:
# Compute all token similarities
scores = query @ keys.T # O(Lk) if using efficient indexing
top_k_indices = topk(scores, k)
attended_values = gather(values, top_k_indices)
output = softmax(scores[top_k_indices]) @ attended_values
RL post-training: Allocate compute budget:
total_training_steps = N
pretrain_steps = 0.9 * N
rl_steps = 0.1 * N
Agentic environment synthesis: Generate diverse tool-use scenarios:
for domain in 1800_domains:
for scenario in [reasoning, planning, execution]:
synthetic_prompt = generate(domain, scenario)
add_to_training_data(synthetic_prompt)
When to Use
- Building state-of-the-art open-source reasoning models
- Applications requiring long-context processing with sparse attention
- Scenarios where tool-use and agentic reasoning are important
- Tasks where post-training compute allocation justifies 10%+ overhead
When NOT to Use
- Real-time inference where sparse attention overhead is prohibitive
- Tasks where full attention is necessary
- Scenarios with limited compute budget for extensive post-training
- Models where tool-use is not required
Key References
- Sparse attention mechanisms and efficient transformers
- GRPO and reinforcement learning for LLM alignment
- Agentic reasoning and tool-use frameworks
- Olympiad-level mathematical reasoning
Related Skills
ADu2021/flow-map-trajectory-tilting
testing
VerifiedTrustedCommunity
Uses flow maps as look-ahead operators to enable principled reward-guided diffusion by predicting trajectory endpoints at any denoising step. Deploy when applying rewards or preferences to diffusion trajectories with meaningful gradients throughout generation.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexible-data-mixture-of-experts
testing
VerifiedTrustedCommunity
Train language models where each expert learns independently on closed datasets, enabling flexible inference with selective data inclusion or exclusion. 41% performance improvement while allowing users to opt out of specific data sources without retraining.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexibility-trap-diffusion-reasoning
data-ai
VerifiedTrustedCommunity
Understand how token generation flexibility in diffusion LMs paradoxically constrains reasoning, as models exploit ordering flexibility to avoid uncertain tokens, and apply simplified approaches that preserve parallel decoding benefits. Use when optimizing diffusion-based language models for reasoning tasks.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flex-continuous-agent-evolution
devops
VerifiedTrustedCommunity
Enable LLM agents to improve continuously during deployment by constructing structured experience libraries through self-reflection on successes and failures—achieving 23% improvement on reasoning without gradient-based parameter updates or external training.
2SKILL.mdUpdated Apr 17, 2026