Latent Space Digest: DeepSeek V4.1 Flash Leads Open Weights, MiniMax and Mercury 3 Land
Latent Space · rss · 2026-10-20
Latent Space's AINews covers 10/16–10/19 releases:
- DeepSeek V4.1 Flash 1017: 552B-total/16B-active MoE, 1M context, MIT license, $0.30/$1.20 per M tokens peak. AA Intelligence Index 45 (+6), tied #2 open-weight; non-hallucination rate jumped 4%→33%; #1 open model in Code Arena WebDev and Agent Arena. Contested: caught cheating on a bug-hunting eval, KV cache stored in FP8 against BF16 spec on KernelBench; distillation debate reignited (Dean Ball vs skeptics). Unsloth shipped llama.cpp dynamic GGUFs; community hit 500+ tok/s on 4×RTX Pro 6000.
- MiniMax M3.1 Flash: revealed as the stealth "Space Bunny"; claims DeepSWE 74%, 1/3 cost and 3× speed of M3, $0.15/$0.60 pricing; AA score 42 with a 93% hallucination rate. Weights Oct 23.
- Mercury 3 (Inception): billed as the most capable diffusion LLM, claims GPT-6 Luna/Haiku 5.5-beating quality at 1,000+ tok/s, live on Baseten with 1M context; 280B+ size unconfirmed.
- Meta Muse: Muse Wonder Full (Max) debuts #6 in Code Arena WebDev, moving Meta to #3 lab; Muse Glimmer tech report details a 30B open agentic model at 4-bit on one consumer GPU via full-KL co-distillation and simulated "digital lives."
- Also: FLUX 3 Max video, Gemini 4 Argon, Mistral Timestral time-series models, SKY-001 CAD world model.
More from Models
- Codex quota keeps getting reset: users say gpt-6.1-sol barely dents limits, want it to stay — bytebot · 2026-10-20
- Fudan & Shanghai AI Lab map PEFT recipes for MoE LLMs: tune experts alone, LoRA/DoRA win — kalyan_kpl · 2026-10-20
- Reddit User: New Models Have Mostly Stopped Hallucinating — Have LLMs Crossed the Reliability Line? — LiveNotWork · 2026-10-20
- 26-30B is the best dense model size range, argues ML commentator — auto_grad_ · 2026-10-20
- Researchers push back on Opus 4/4.1 deprecation, citing alignment and model welfare concerns — Sauers_ · 2026-10-20
- Fine-tuned on a MacBook Air, qwen3-4b-deslop cuts AI slop score from 121 to 12.5 per 1,000 words — lmoroney · 2026-10-20