Skyfall AI released SKY-001, a world model that designs physical products by planning parts and build steps in 3D, matching GPT-6 Astra's CAD accuracy at roughly 150x lower cost per design.
Levanto Labs introduces Sage2, a streaming decision model that computes each token of a conversation only once, reusing cached context so each new message adds only its own token cost. It delivers typed answers with calibrated probabilities and claims 3x faster decisions on progressive context, plus extraction that is 166x cheaper than Opus 5.5.
Google's team used Gemini 4 Argon within Antigravity to script 3D Blender boot animations and SuperCollider audio, rendering a video displayed on a vintage CRT TV via Raspberry Pi.
The organization introduces Hydrogen-1.0, an open-weight model fine-tuned from GLM-5.3 on internal scientific data, claiming 38.3% on FrontierScience-Research (vs GPT-6-Astra's 33.3%) and 93.0% on GPQA-Diamonds, in service of open scientific discovery.
Moonshot AI announced Kimi K3.1 Pro, its most capable 3T-class open-weight MoE model with 1M-token context, and Modal will offer day 0 support for it this week.
Unsloth published benchmark results for its DeepSeek Dynamic V3.1 GGUF quants, claiming >10–20% gains over GSQ-RCO on agentic benchmarks like Terminal Bench and DeepSWE at 512K context, and statistically disputing GSQ-RCO's quantization recovery claims.
AtomicChat released GSQ-RCO GGUF quantizations of DeepSeek-V4.1-Flash-1017 with refusal directions projected out of the dense residual writers, letting a 763B MoE model run on a single 12 GB GPU via Strata with minimal MMLU degradation.
LongCat-2.5 is introduced as a natively multimodal model designed for agentic tasks, combining vision-language capabilities with tool-use and long-horizon task performance.
ISTA DASLab released a Coder build of DeepSeek-V4.1-Flash-1017, a 763B sparse MoE model compressed to roughly 1 bit per transformer parameter via GSQ quantization and RCO-based expert pruning (192 of 384 experts kept), retaining 98.7% of LiveCodeBench v6 and 94.8% of SWE-bench Verified performance in a 72.1 GB GGUF.
Egzakta AI Lab released KVARK Sparkle 1.0, an open-weight 27B model based on Qwen3.8-27B that cuts thinking tokens by up to 67% and doubles median response speed while maintaining comparable accuracy across five standard benchmarks.
A user shares an appreciation post for the DeepSeek V4.1 Flash 1017 iteration, praising its speed, low cost, and reliability for everyday tasks and claiming it finally makes closed-source models unnecessary for them.
Fovea-4B is an open-weight 4B vision model that brings Jev-style typed decision contracts to images, leading ImageJev-style tasks with 80.6 chance-corrected accuracy and strong calibration (ECE 0.035), released under Apache-2.0.
用户分享观点,认为 Claude Opus 5.5 的 High 模式优于 XHigh 模式。
JetBrains released Mellum2.1-12B-A2.5B-Thinking-MTP, a 12B MoE (2.5B active) thinking coding model with a multi-token prediction head for speculative decoding in vLLM. The update focuses mainly on reinforcement learning post-training in real repository environments to improve agentic software engineering tasks.
Anthropic appears to be gearing up for the launch of Mythos 5.5, with signs that its product webpage is already being prepared ahead of the release.
humans& releases &Tutor, an AI model trained to teach by training against Persimmon as a simulated learner, achieving state-of-the-art results on the TutorMoments benchmark and outperforming frontier models like Claude Sonnet 5.5 and GPT-6 Astra under plain prompting.
OpenAI发布 GPT-Rosalind-6 系统卡增补,说明该模型复用 GPT-6 Astra 权重,并在 Genomics 与生物化学能力/安全评测中报告了相较于 GPT-Rosalind-5.5 的显著提升及相应的防护措施。
A 4B open-weight coding model (Qwen3.5-4B-GLB-3D-Coder) that generates Three.js scenes with GLB export for creating 3D game assets usable in engines like Godot is being released as an initial checkpoint with full weights and training data to follow.
An updated 2.3bpw quantized checkpoint of a Qwen 27B model for 12 GB Ampere cards ties Qwen3.6 BF16 on LiveCodeBench (84.3% vs 83.9%) and cuts KLD error by ~40% versus the previous release, using the new AmperInteger+ quant format in llamAmpere.
Silencio-2 is an open-weight 18M-parameter TTS fine-tune of Kyutai's Pocket TTS that follows inline emotion and sound tags (laugh, sigh, whisper, angry, etc.) and runs ~20x realtime on CPU. The author shares training details and lessons, notably that an AST/AudioSet clip-level filter silently dropped rare sound events, so it was removed entirely.