GLM Scale OpenGL YouTube - Search Videos

Patronus AI on X: "In DeepSeek, GLM and Kimi, each newer release uses fewer filler words. DeepSeek 26.6%→13.3%, GLM 18.9%→13.6%, Kimi 18.0%→9.8%. Our tasks include 8 from SWE-bench Verified, 5 from SWE-bench Pro and 7 from Terminal-Bench 2.1, all in a terminal agent at high reasoning effort." / X

Patronus AI on X: "In DeepSeek, GLM and Kimi, each newer release uses fewer filler words. DeepSeek 26.6%→13.3%, GLM 18.9%→13.6%, Kimi 18.0%→9.8%. Our tasks include 8 from SWE-bench Verified, 5 from SWE-bench Pro and 7 from Terminal-Bench 2.1, all in a terminal agent at high reasoning effort." / X

Lilac on X: "OMG Kimi / DeepSeek / GLM free on NVIDIA Build OpenAI-compatible free endpoints. what people are using: kimi k3.1 pro (1M context where listed) deepseek v4.1 flash glm 5.3 flash how: 1. https://t.co/9SSZ3kAvao and sign up (NVIDIA Developer) 2. generate a free API key 3. base URL: https://t.co/a6UGU7wI03 4. call with the model id from the model page (e.g. deepseek-ai/deepseek-v4.1-flash, z-ai/glm-5.3-flash) enjoy." / X

Lilac on X: "OMG Kimi / DeepSeek / GLM free on NVIDIA Build OpenAI-compatible free endpoints. what people are using: kimi k3.1 pro (1M context where listed) deepseek v4.1 flash glm 5.3 flash how: 1. https://t.co/9SSZ3kAvao and sign up (NVIDIA Developer) 2. generate a free API key 3. base URL: https://t.co/a6UGU7wI03 4. call with the model id from the model page (e.g. deepseek-ai/deepseek-v4.1-flash, z-ai/glm-5.3-flash) enjoy." / X

Niklas on X: "Two LLMs go to war on equal hardware: GLM-5.3-Flash vs DeepSeek-V4.1-Flash, 2 DGX Sparks each. 2x DGX Spark, 256 GB: GLM-5.3-Flash Mia's TensorFold recipe, commit bf5f945 Mia's EXL3 quant, 4 bpw 2x DGX Spark, 256 GB: DeepSeek-V4.1-Flash Mia's TensorFold recipe, commit 0e7d421 Mia's EXL3 quant, 2.9 bpw Median thinking time per decision: GLM 173.5 s DeepSeek 33.0 s Decisions in 17 mins of game time: GLM 11 DeepSeek 44 The game does not wait for a model, so a faster thinker gets more

Niklas on X: "Two LLMs go to war on equal hardware: GLM-5.3-Flash vs DeepSeek-V4.1-Flash, 2 DGX Sparks each. 2x DGX Spark, 256 GB: GLM-5.3-Flash Mia's TensorFold recipe, commit bf5f945 Mia's EXL3 quant, 4 bpw 2x DGX Spark, 256 GB: DeepSeek-V4.1-Flash Mia's TensorFold recipe, commit 0e7d421 Mia's EXL3 quant, 2.9 bpw Median thinking time per decision: GLM 173.5 s DeepSeek 33.0 s Decisions in 17 mins of game time: GLM 11 DeepSeek 44 The game does not wait for a model, so a faster thinker gets more

Adam Holter on X: "You would think that there's a trade-off between speed and accuracy when it comes to providers, but from Artificial Analysis results from GLM 5.3 Flash, the fastest providers also had very strong quality It was the slower providers that tended to drop off" / X

Adam Holter on X: "You would think that there's a trade-off between speed and accuracy when it comes to providers, but from Artificial Analysis results from GLM 5.3 Flash, the fastest providers also had very strong quality It was the slower providers that tended to drop off" / X

Niklas on X: "Most requested matchup yet: @MiaAI_lab's new GLM-5.3-Flash recipe on 2 DGX Sparks vs DeepSeek-V4.1-Flash on a Mac Studio First time we see the game with another skin! It tests Mia's new GLM recipe, version 1104, against Bob's DeepSeek recipe, 256 GB each. Only one base survives. 2x DGX Spark, 256 GB: GLM-5.3-Flash Mia's TensorFold recipe, version 1104 Mia's EXL3 quant, 4 bpw 1x Mac Studio M5 Ultra, 256 GB: DeepSeek-V4.1-Flash Bob's TensorFold Metal recipe Bob's quant, experts at 2

Niklas on X: "Most requested matchup yet: @MiaAI_lab's new GLM-5.3-Flash recipe on 2 DGX Sparks vs DeepSeek-V4.1-Flash on a Mac Studio First time we see the game with another skin! It tests Mia's new GLM recipe, version 1104, against Bob's DeepSeek recipe, 256 GB each. Only one base survives. 2x DGX Spark, 256 GB: GLM-5.3-Flash Mia's TensorFold recipe, version 1104 Mia's EXL3 quant, 4 bpw 1x Mac Studio M5 Ultra, 256 GB: DeepSeek-V4.1-Flash Bob's TensorFold Metal recipe Bob's quant, experts at 2

Adam Holter on X: "For some models like GLM 5.3, you have to be careful what provider you're picking But DeepSeek for some reason seems to be really consistent" / X

Adam Holter on X: "For some models like GLM 5.3, you have to be careful what provider you're picking But DeepSeek for some reason seems to be really consistent" / X

Surya on X: "Inkling 2 Small can paint 👩‍🎨!! (Video + Audio made in code) Comparison below between GLM 5.3 Flash, DeepSeek V4.1 Flash and Qwen 3.8 Flash on the same prompts. ⬇️" / X

Surya on X: "Inkling 2 Small can paint 👩‍🎨!! (Video + Audio made in code) Comparison below between GLM 5.3 Flash, DeepSeek V4.1 Flash and Qwen 3.8 Flash on the same prompts. ⬇️" / X

Adam Holter on X: "Mistral may have trash models, but Vibe subsidizes GLM 5.3 way better than https://t.co/9nfFlmurjg does" / X

Killy on X: "Model offloading visualized GLM 5.3 edition GLM-5.3 is a 744B moe model, with 13 GB read for every token. Another 318 GB is 19,200 experts, and one token uses 8 of 256 in each layer, about 3%. So we put that 13 GB on GPUs and experts live in RAM, which leaves lots of VRAM open which is exactly what we want Leftover VRAM is used to hold second copies of experts used most often by sorting the bytes by how often they are read. One byte on GPU reads roughly x3-5 faster than CPU reading

Md Ismail Šojal 🕷️ on X: "Running GLM-5.3-Flash locally on 2x RTX PRO 6000s - 200 tokens/sec and watching it generate long, detailed analysis Cloud pricing makes this look expensive." / X

Mia on X: "MiniMax-M3.1-Flash vs GLM-5.3-Flash In a strategy game! Great and original concept!!" / X

noclipepe on X: "How LongCat-2.5 compared with Gemini, GPT and GLM on the same SVG task One shot. Same prompt: an animated SVG of a pelican riding a bicycle. LongCat-2.5 · $0.25 The most complete scene. Fish in the pouch, a ringing bell, a passing seagull, parallax background. Legs follow the pedals. Gemini 4 Argon · $0.37 Best colors and a bold, clean pelican. The background is simpler. GLM-5.3 Flash · $0.24 Clean and accurate drawing, a calm seaside scene. Fewer details and less motion. GPT-6.

Wayne Lowry on X: "The debug line says ground true. The hero is floating in mid-air GLM 5.3 Flash, one prompt, one hour. It ends on the game's own GAME OVER screen Would you have trusted that debug line?" / X

Hardeep on X: "Same AI model. 30% or 81%. The harness decided. IBM Research gave 3 agent harnesses the same open model (GLM-5.3-flash) and the same 90 AppWorld tasks. Scores ran from 30% to 81%. 1. A harness is the code around the model. It runs the loop, calls tools and checks the results. 2. IBM's CUGA passed 73 of 90 tasks. Hermes passed 42. LangChain's Deep Agents passed 27. 3. CUGA's edge is plain engineering. It splits the task one app at a time, shows each step only that app's tools, and

ＺＥＦＩ on X: "nvidia is hosting frontier AI MODELS FOR FREE no card for the trial key: > nvidia nim: deepseek v4.1 flash, glm 5.3 flash, mistral large 4, step 5 preview. openai-compatible base url: https://t.co/8CvetTVutj > openrouter free right now: gemma 4 31b + 26b, flint, xing 4.0, north mini code, dots3-note, lfm 2.5: https://t.co/0BPng3WnQQ > opencode zen $0: deepseek v4.1 flash, glm 5.3 flash, mimo v2.6 flash, flint 40b, beam: https://t.co/zDdZJXnFb5 > orcarouter free: deepseek v4.1 flash +

Niklas on X: "Two AI models just crashed their armies into each other. Tomorrow 8 AIs fight in my tournament. The big fight from my strategy game, called by two AI casters. Mia's GLM-5.3-Flash recipe on 2 DGX Sparks vs Bob's DeepSeek-V4.1-Flash recipe on a Mac Studio. In this clip: Big fight starts at 14:45 of game time DeepSeek's army value at 15:18: 1560 2x DGX Spark, 256 GB: GLM-5.3-Flash Mia's TensorFold recipe, version 1104 Mia's EXL3 quant, 4 bpw 1x Mac Studio M5 Ultra, 256 GB: DeepSeek-V4

S 🫧 (@dyfwsammy) on X

3.3K views5 days ago

Stanislav Sorokin on X: "OpenRouter says Auto Code costs 74% less than Opus 5.5. On our 33 coding runs it saved 28%. OpenRouter shipped Auto Code on 5 Nov: a router for coding agents that "benchmarks similar to Opus 5.5 while costing 74% less." We gave both the same 11 coding tasks with hidden tests today, 3 runs each. → Pass rate: Auto Code 33/33, Opus 5.5 33/33 → Cost: $0.339 vs $0.473, 28% less → Median answer time: 7.6 s vs 5.7 s → Routed to: GPT-6.1 Sol 21 times, GLM-5.3 12 times, Claude 0

Laz252 on X: "You can now run Claude Code for completely free, with almost unlimited usage. Someone on GitHub just built an Al tool that gives you over 7 billion free Al tokens every single month. And that includes models from ChatGPT, Gemini, GLM, Kimi, Grok, Mistral, Qwen, and much more. It's called FreeLLMAPI. It links your Claude Code to over 600 free Al models, all through one single setup. So the moment your Claude limit runs out, it automatically switches you to the next best model. Your

Ozan (@OzanTRAI) on X

297 views4 days ago

See more