🎁 Follow & DM to claim free compute pack ! An all-in-one hub for datasets, tutorials, papers and benchmarks. Join our community⬇️ https://xitter.cf/t.co/bN33gcLbB7
Joined March 2019
- Tweets2.1K
- Following159
- Followers201
- Likes755
Pinned Tweet
📢 Regarding HyperAI Account Actions
Recently, HyperAI detected a number of accounts running malicious and attack-oriented code, alongside large-scale registrations and abuse of promotional resources. These activities have negatively impacted cluster stability, platform security, and the experience of legitimate users.
To ensure normal platform operations, we are conducting large-scale cleanup and suspension actions against related accounts, while further strengthening our risk-control and review systems.
During this process, a small number of legitimate users may be affected by mistake, and we sincerely apologize for any inconvenience. If your account has been impacted, please contact us via our official email ([email protected]) or through Discord. Our technical team will manually review your case and restore access as soon as possible.
Finally, we sincerely appreciate the community’s understanding and support for our ongoing platform security efforts.
— HyperAI Team
May 24, 2026
Gregor's head — human or beetle?
Four models, one identical prompt:
Seedream 5.0 / GPT Image 2.5 Sunburst / FLUX 3 Image Klein / Grok Imagine 2.0
FLUX 3 Image Klein is the only open-weight model in the lineup — 5.6B parameters, up to 10 reference images — and it does not fall far behind.
Which one matches the creature you pictured while reading?
#Kafka #TheMetamorphosis #AIImage #FLUX #Seedream #Grok #OpenSource
Nanbeige4.3-33B vs Qwen3.8-27B
Measured on a single A100 80GB, both native BF16, no 4-bit or 8-bit quantization.
= = =
MODEL
Released Oct 24, 2026 by Nanbeige LLM Lab
Apache 2.0, 33.1B dense, 524,288-token context, text only
BENCHMARKS
Terminal-Bench 2.1
Nanbeige 80.9 / Qwen 73.0
DeepSWE
Nanbeige 65.5 / Qwen 42.5
SWE-Bench Pro
Nanbeige 60.3 / Qwen 61.7
GDPval
Nanbeige 84.8 / Qwen 86.4
All figures are vendor-reported, harnesses chosen by the vendor, no independent reproduction yet.
Muse-Glimmer-30B on Terminal-Bench 2.1: 61.8 here, 51.7 as published by Meta.
Gemma4-31B on GPQA Diamond: 91.3 here, 84.3 as published by Google.
MEMORY
52 layers, 8 KV heads, head dim 128, BF16
208 KiB of KV cache per token
512K context maps to 104 GiB of cache, against 61.69 GiB of weights
On the same A100 80GB:
Resident after load: 74,265 MiB vs about 62,000 MiB
Available KV space: 9.41 GiB vs 18.95 GiB
Max context: about 47,440 tokens vs 150,000+
A 16K request supports roughly 2.9 concurrent sequences. A 64K request exceeds one card.
Both the vLLM and SGLang recipes in the model card specify four-way tensor parallelism.
FIVE ENGINEERING TASKS
Ledger audit: 4/4 in 48.47s / JSON truncated at line 13
SRE tool calling: 6/6 / 6/6
Concurrent singleflight: 0/6 / 0/6, both ran out of 4,096 tokens inside the thinking block
Multi-defect diagnosis: neither produced complete code within 4,096 tokens
DEPLOYMENT
LlamaForCausalLM architecture. vLLM loads it through the native Llama path; reasoning and tool-call parsers reuse the Qwen3 conventions.
Official release is BF16 only. GGUF, FP8, NVFP4, MLX and EXL3 are community uploads.
SGLang supports FP8 KV cache and n-gram speculative decoding.
A DGX Spark deployment uses Nanbeige's own 3B model as a drafter, raising single-request throughput from about 7 to 12.6 tok/s.
MODALITY
Text only. The chat template returns a notice on image input.
Qwen3.8-27B ships a 27-layer vision encoder with image and long-video support.
#LLM #qwen #Nanbeige #A100 #LocalLLM #vLLM #SGLang #GPU #Benchmark #OpenSource
I just published Nanbeige4.3–33B: Deployment and Evaluation medium.com/@hyperai/nanbeige…
🚀 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
GitHub: github.com/hi3d-official/Hi3…
Open source: hi3d.ai/hi3d-1-0-open-source
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
C̶̨̛̦̪̰̘̙̦̓̈́l̸̢̧̛̺͈̫̈̂ȩ̴̛̛̘̦̓f̶̨̢̛̛̦̪̰̘̙̦̓̈́-̸̢̧̛̺͈̫̈̂F̶̨̛̦̪̰̘̙̦̓̈́l̸̢̧̛̺͈̫̈̂ȃ̴̛̛̘̦̓ș̶̨̢̛̛̪̰̘̙̦̓̈́ḧ̸̢̧̛̺͈̫̂
typed probabilistic decision model — live on HyperAI 👇
hyper.ai/notebooks/clef-flas…
DeepSeek closes $15B funding round led by Tencent and CATL, per Bloomberg.
• 100B yuan, one of China's largest AI fundraises
• Upsized twice: 50B → 80B → 100B yuan
• IPO preparation underway (STAR Market)
• Still trails OpenAI/Anthropic ($100B+ each) but closing the gap
The open-source champion is going public.
#DeepSeek #AI #IPO #Tencent #CATL #OpenSource
