Alibaba's Qwen team released Qwen3.8-Flash-Next on August 26, 2026, an open-weight experimental model with 125B total parameters but only 6B active per token, previewing the architecture intended for …
DeepSeek released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family, adding visual modules to the existing DeepSeek-V4-Flash architecture and scoring 36.5 Pass@1 o…
Z.ai's GLM 5.3 Flash ranked first in a 12-task Editorial Craft benchmark with a mean score of 0.94 and an estimated cost of $0.0002 per task, according to RuntimeWire. Step-3.7-flash followed at 0.92,…
Alibaba released Qwen3.8-Flash-Next, an open-weight multimodal model with a 125B-parameter Mixture-of-Experts backbone plus 51B n-gram embedding parameters and 6B active parameters per token, which ou…
Ornith AI released Ornith-1.5-397B, a 403B-parameter mixture-of-experts agentic-coding model, on 2026-08-18, claiming coding scores on par with Claude Opus 4.8 and ahead of GLM-5.2 and DeepSeek-V4-Fla…
A homelab user deploying DeepSeek-V4-Flash-0731 on two DGX Spark units achieved 1M token context with an NVFP4 KV cache using a custom vLLM fork, reporting a KV cache size of 1,492,347 tokens and maxi…
DeepSeek released deepseek-v4-flash-vision-exp, an experimental multimodal variant of its V4-Flash model, on August 21, 2026, matching text-only capabilities while accepting images at the same per-tok…
DeepReinforce released Ornith-1.5, a family of open-source AI models with 9B, 35B, and 397B parameters, under the MIT license, claiming the flagship outperforms Claude Opus 4.8 on Terminal-Bench 2.1 (…
Ornith AI released Ornith-1.5, a family of open-source language models under the MIT License, featuring a self-improvement training loop that generates tasks and improves policy through reinforcement …
Ornith-1.5, a new family of foundation models from the Ornith project, achieves state-of-the-art performance among open-source models of comparable size, with the flagship Ornith-1.5-397B scoring 86.1…
DeepReinforce released Ornith-1.5, a family of open models in 397B, 35B, and 9B sizes, extending the self-scaffolding framework into a closed self-improvement loop where the system proposes tasks, gen…
A developer reports achieving 99.4 tok/s prompt processing and 10.1 tok/s text generation with DeepSeek-V4-Flash-0731 (UD-Q4_K_XL GGUF) on four NVIDIA RTX 3060 12GB cards using llama.cpp build b10181,…
RunInfra lists DeepSeek V4 Flash, an LLM served as deepseek-ai/DeepSeek-V4-Flash-0731, at $0.13 per 1M input tokens and $0.27 per 1M output tokens, with a 1,048,576-token context window and OpenAI-com…
A developer at Petabit Scale published a portable reproducer showing that DeepSeek-V4-Flash-0731 collapses into repeating the short action-announcement "Run." instead of emitting tool calls during lon…
DeepSeek released DeepSeek V4 Pro 0813 on OpenRouter, following the open-weights releases of DeepSeek-V4-Pro in April and DeepSeek-V4-Flash-0731 in July, though open weights for the new version are no…
DeepSeekV4SSD, an experimental app from developer yanun0323, streams routed experts from SSD to run all 284 billion parameters of DeepSeek-V4-Flash-0731 on an M-series Mac with about 30 GB of memory, …
Alibaba's Qwen 3.8-Max, marketed as outperforming GPT-5.6, Sol Max, and Fable 5 on agentic computer use, ranked mid-pack or last in independent VulcanBench tests, a discrepancy driven by time budgets:…
DeepSeek released DeepSeek-V4-Flash-0731 on July 31, an updated version of its existing deepseek-v4-flash API model that now outperforms DeepSeek's V4-Pro-Preview on all nine agent and coding benchmar…
DeepSeek released DeepSeek-V4-Flash-0731 on July 31, a $0.14 per million input tokens model that outperforms its own flagship, a 1.6 trillion parameter model, on all nine agent benchmarks published by…
A single AMD MI300X GPU with 192 GB of HBM3 now serves DeepSeek's 284B-parameter DeepSeek-V4-Flash-0731 checkpoint in mixed FP4+FP8 format, requiring nine patch overlays against a vLLM ROCm nightly pl…