🚀 FLUX 3 Image Klein is now live on HyperAI • 5.6B params — smallest in its class • Native 4K generation + pixel-perfect editing • Multi-reference composition (up to 10 images) • Outperforms larger models on text-to-image + editing No API needed — one-click on HyperAI 👇 hyper.ai/notebooks/flux-3-im… #FLUX3 #BlackForestLabs #TextToImage #ImageEditing #OpenWeights #AIart #HyperAI

Fresh datasets on HyperAI this week 🧮 Math & Code: • Orca-Math — SLMs solving grade school math (Microsoft) • OpenMathInstruct-1 — math instruction tuning (NVIDIA) • OpenCodeInstruct — code instruction tuning (NVIDIA) • OpenMath GSM8K / MATH — masked mathematical reasoning (NVIDIA) 🤖 Agents & Benchmarks: • TaskBench — task automation benchmark (Microsoft) • SCBench — long-context benchmark (Microsoft) • APEX-Agents — agent benchmark • DABstep — multi-step reasoning for data analysis • ProfBench — professional skills benchmark (NVIDIA) • RPR — Reasonable Preference Reversal for alignment (Microsoft) 🦾 Physical AI & Robotics (NVIDIA): • Nemotron VLM v2 — vision-language model data • SimReady Warehouse OpenUSD assets • Robotics Manipulation Kitchen Demos • Datacenter CFD — physics simulation 🎙️ Speech & Language: • NOTSOFAR-1 — far-field meeting transcription (Microsoft) • Last Translation Benchmark Link in comments 👇

Hi3D just open-sourced their V3.0 3D generation model. 2048³ ultra-high-precision, commercially ready, fully self-developed in China. More info👇 #AI3D #3Dgeneration #opensource

Every documented way to run DeepSeek-V4.1-Flash-1017 locally, with community-measured data: Single RTX 5090 (32GB) + 128GB RAM: • acrogenesis recipe-kit: 21.4 tok/s fresh, 25 tok/s cached. IQ2_XXS quant, 359GB. Perplexity 7.57 vs FP8's 6.34. • dsx: 12-15.2 tok/s. 2-bit quant, 148GB + 203GB Engram. KL 0.395, top-1 74.1% vs FP8. • JigSawPT llama.cpp fork: 5.1 tok/s fresh, 21.3 tok/s resident. MXFP4, 502GB. 0.9967 logit correlation with reference. Single RTX 4090 (24GB) + 90GB RAM: • Xolotl (Strata fork): ~27 tok/s. GSQ mixed ternary/2-bit/3-bit, 152GB. Hardware floor for this model. Dual RTX 3090 (48GB total) + 128GB RAM: • dsx: 28.9-37.2 tok/s decode, 301-508 tok/s prefill. The VRAM jump from 32→48GB holds ~3,300 more experts resident. Dual RTX 3090 Ti + NVLink + 512GB RAM: • fastllm (igerry fork): 9-20 tok/s. Ran 13-hour agent tasks without crashing. Tokenizer, image input, and harness fixes all upstreamed. Mac Studio M5 Ultra 256GB: • oMLX LSQ3-G128: 18.8 tok/s. MMLU-Pro 51.5, GPQA 16.2, LiveCodeBench 62.5. Only Mac build with published evals. Dual DGX Spark (256GB unified): • TensorFold + EXL3 2.9bpw: 74.9 tok/s single-stream, 130-155 tok/s 4-stream. Prefill 2,000-2,183 tok/s. 524K needle-in-haystack 100% accurate. Zig runtime, zero Python. 8× B200 (production): • vLLM aggregated: 990.6 tok/s/GPU system output. Disaggregated: 1,087.7. 830K tokens/s/GPU at 100 tok/s/user. 1.3B tokens per dollar TCO. 2× GH200: • SGLang outperforms vLLM by 17.6% on 8K prefill (10,262 vs 8,725 tok/s), 18.2% on code review. 16× Ascend 910B3 NPU (2× Atlas 800 A2): • vllm-ascend W8A8: 222.1 tok/s at 16 concurrent. 0.00% error rate across 98 requests. 1M context verified at 1,048,576 tokens. In production. 15+ quantization formats documented: GGUF (Q4/Q8/IQ1/IQ3), EXL3 (2.9/3.5bpw), MLX (LSQ3/oQ3e/FP8/4bit), dsx 2-bit, GSQ, MXFP4, and more. Every number above comes from a published repo or benchmark. Full configs and links in the complete guide: #DeepSeek #OpenSource #LLM #LocalAI #GPU #MachineLearning #AI

Case: Ran DeepSeek V4.1 Flash 1017 on a single RTX 5090 #deepseek #LocalLLM

DeepSeek just open-sourced their ENTIRE production inference stack 🏭 The whole factory: 🔥 DeepInfer — production inference engine (Apache 2.0) • 25K tok/s/GPU decode on 64 GPUs • DSpark speculative decoding • NVIDIA SM100 + Ascend 950 NPU 🛠️ deepseek-recipe — official tokenization matched to API behavior Plus the stack they already dropped: ⚡ FlashMLA, DeepGEMM, DeepEP, 3FS, DeepSeek Harness, Engram #OpenSourceAI #DeepSeek #LLM #AIInfrastructure

A weaker model with a better harness beat a stronger model with its original harness. GLM-5 + Harness²: 75.7 Gemini 3.5 Flash as shipped: 72.4 WHAT'S HARNESS² The model is a chef. The skill is trained in. You can't change it. The harness is the kitchen around him: where the recipes sit, how the knives are laid out, the prep list, how the sous-chefs coordinate, what gets cooked first. Same chef. Different kitchen. Very different dinner service. That's the 75.7 vs 72.4. A slightly worse chef in a well-organized kitchen beats a better chef in a messy one. Until now, a human stood next to the chef, watched him fail, and rearranged the kitchen. Slow, expensive. Harness² lets the chef rearrange it himself, mid-service. The hard part: a new dish has no answer key. How do you know the new layout is better? So it cooks the same dish twice, two deliberately different layouts, and compares the process. Not "does it taste right." "Did it stop getting stuck." The square in the name is the second level: after service, it writes down "knives on the left worked better," and reads that note before the next one. Rearranging the kitchen is one level. Rearranging how you rearrange it is the second. The counterintuitive part. Let the chef rearrange freely, no comparison, no notes, and the food gets worse. 73.4 → 63.2. The point isn't that it edits itself. It's that it compares before it edits. #harness #Google