Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official1
static.stepfun.com2026-02-02officialHN front pageHN 231 pts · 91 comments3,135 words
Step 3.5 Flash is a sparse MoE open-source model with 196B total parameters that activates only 11B per token. It achieves 100–300 tok/s throughput using MTP-3 and scores 74.4% on SWE-bench Verified. The model supports a 256K context window via a 3:1 Sliding Window Attention ratio and runs securely on high-end consumer hardware like Mac Studio M4 Max. It distinguishes itself through a unique "Think-and-Act" synergy in tool environments.written, reviewed, cited
- The model is compatible with Claude Code, serving as an efficient backend for agent-led development.
- Step 3.5 Flash synthesizes a comprehensive research report of approximately 10,000 words on early childhood science education.
- The model achieves a score of 65.27% on the Scale AI Research Rubrics benchmark for deep research tasks.
Analysis8
arxiv.orgauthority24,483 words
Step 3.5 Flash is a sparse Mixture-of-Experts model with 196B total parameters and 11B active parameters designed for efficient agentic inference. It achieves 85.4% on IMO-AnswerBench and 86.4% on LiveCodeBench-v6 while utilizing Metropolis Independence Sampling-Filtered Policy Optimization (MIS-PO) for stable reinforcement learning. The architecture scales the context window to 128k and employs interleaved 3:1 Sliding Window/Full Attention to minimize latency during multi-round interactions.written, reviewed, cited
- Step 3.5 Flash Base scores 31.6% on SimpleQA, surpassing DeepSeek-V3.2-Exp Base despite using only one-third of the parameters.
- The architecture is constrained to under 200B parameters to enable high-performance inference within the 128GB memory budget of high-end workstations.
arxiv.org2026-02-12authorityHN 1 pts · 1 comments390 words
Step 3.5 Flash is a sparse Mixture-of-Experts model that bridges frontier-level agentic intelligence and computational efficiency. It pairs a 196B-parameter foundation with 11B active parameters and uses interleaved 3:1 sliding-window/full attention with Multi-Token Prediction (MTP-3) to reduce latency. The model achieves 85.4% on IMO-AnswerBench and 86.4% on LiveCodeBench-v6, demonstrating performance comparable to frontier models like GPT-5.2 xHigh.written, reviewed, cited
github.comauthority2,404 words
Step 3.5 Flash is a 196B parameter sparse MoE model that activates only 11B parameters per token using a strategy of selecting Top-8 experts from 288 routed experts per layer. It achieves 100–300 tok/s throughput via 3-way Multi-Token Prediction and scores 74.4% on SWE-bench Verified while supporting a 256K context window with 3:1 Sliding Window Attention. The model is available via OpenRouter free and paid tiers, and the China platform requires +86 phone verification.written, reviewed, cited
- The model requires a specific patch file to run on vLLM until full MTP3 support is integrated.
- Deployment on llama.cpp mainline may require building with PR #1867 to resolve tool calling issues.
github.com2026-02-05authorityHN 1 pts · 0 comments
docs.api.nvidia.com1,393 words
computertech.co2026-02-192,744 words
Community1
old.reddit.com2026-02-02HN 2 pts · 1 comments
