deepseek v4 1 flash 552b moe model hugging face a deepseek v4.1 flash stylised whale v2

DeepSeek V4.1 Flash 552B MoE Model Released on Hugging Face. It Beats V4 Pro on Agents and Trails It on Knowledge

September 10, 2026
Artificial Intelligence
Comment off

DeepSeek released V4.1 Flash on Hugging Face on 10 September 2026: a 552B mixture-of-experts model under the MIT licence that activates 8B parameters to read and 16B to write, stores its global KV cache in 890 bytes per token, and replaces V4 Pro for every API caller from 04:00 UTC on 14 September. Its own model card shows it leading V4 Pro on all 12 agentic rows and trailing it on 12 of 16 base-model rows, including a 12.9-point gap on SimpleQA-Verified. This breakdown covers the causal encoder-decoder architecture, the KV cache arithmetic, the new price sheet with a worked agent bill, the 511 GB checkpoint and 614 GB serving floor, the mismatched reasoning-effort aliases, and a pre-cutover checklist.

Tencent Releases Hy4 preview: An Open-Weight MoE Model With a 1M Context Window

August 28, 2026
Artificial Intelligence
Comment off

Tencent open-sourced Hy4 preview on 28 August 2026 under Apache 2.0: a 770B Mixture-of-Experts model that activates just 49B parameters per token and reads a one-million-token context window. This breakdown covers the full architecture, from 256 routed experts per layer to Gated DeepSeek Sparse Attention and the built-in speculative decoding layer; every benchmark figure Tencent published, including the 163-expert blind evaluation against GLM 5.3 and Kimi K3; the API price list against GPT-5.6 Sol; the eight-GPU serving recipes; and the four caveats worth naming before any of it reaches production.

Alibaba Releases Qwen3.8-Flash: A Multimodal 125B MoE Model That Previews Qwen4

August 26, 2026
Artificial Intelligence
Comment off

Alibaba open-weighted Qwen3.8-Flash-Next on 26 August 2026: a multimodal mixture-of-experts model with 125 billion parameters, a separate 51-billion-parameter N-gram embedding table, and just 6 billion parameters activated per token. This breakdown covers the four rebuilt subsystems — Gated DeltaNet paired with Qwen Sparse Attention at block granularity, a Gated Residual stream widened to four gated branches, the N-gram table that offloads to host RAM, and the Muon plus AdamW training recipe with batch-size warmup removed — alongside the 48-layer stack of 512 experts that fires eleven per token, the published benchmark table showing 62.5 on SWE-bench Pro against 53.4 for Claude Opus 4.6 and 84.5 on AndroidWorld against 62.0, the single loss on Humanity’s Last Exam at 35.9 against 40.0, the unverifiable one-ninth training cost claim, the 262,144-token native context extended to a million with YaRN, hosted pricing of $0.16 and $0.47 per million tokens against $2.00 and $6.00 for Qwen3.8-Max, the real hardware bill from a 172.78 GiB FP8 checkpoint down to a 111 GB four-bit GGUF, the qwen-community-1.0 licence that is not Apache 2.0, and a buyer’s checklist for treating a preview checkpoint as a production dependency.