{"full_name":"Niko1221/Strata","owner":"Niko1221","name":"Strata","url":"https://github.com/Niko1221/Strata","description":"Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.","homepage":"","topics":[],"language":"C++","license":"MIT","avatar":"https://avatars.githubusercontent.com/u/86886444?u=0252622cf8ec18e10cf16f39f288c16035be1251&v=4","created_at":"2026-09-24T16:40:20Z","pushed_at":"2026-11-10T08:34:32Z","metrics":{"stars":42671,"forks":4664,"watchers":202,"issues":2702,"prs":2881,"releases":66,"issues_open":null,"issues_closed":null,"prs_open":null,"prs_closed":null,"prs_merged":null,"discussions":0},"growth":{"day":{"stars":444,"forks":47,"base_day":"2026-11-09"}},"history":{"dates":["2026-11-09","2026-11-10"],"stars":[42227,42671],"forks":[4617,4664],"watchers":[202,202],"issues_open":[null,null],"prs_open":[null,null]},"boards":{"daily":82,"daily_rate":95,"new":6},"same_owner":[],"issues":[{"number":3330,"title":"[Other] Opinion: Strata app should be separate repo from Strata Engine","comments":51,"created_at":"2026-10-20T03:11:36Z"},{"number":2197,"title":"[Engine] Turing (sm_75): make the STRATA_SM75_PERF series the default, with expert mode 8 for i-quant packs","comments":47,"created_at":"2026-10-13T00:14:36Z"},{"number":3976,"title":"[Engine] Layer-major prefill on a CUDA layer split: 32K prompts 2.0x faster, 0.47x drive reads (Qwen3.8-Flash-Next, v0.2.4.2)","comments":28,"created_at":"2026-10-24T15:08:10Z"},{"number":4841,"title":"[Engine] Performance regression with engine 0.1.54/55 on 3090","comments":25,"created_at":"2026-10-31T18:01:04Z"},{"number":4610,"title":"[Benchmark] record the host side in the standard report: measured PCIe per card, NUMA, memlock, GPU temperature and power","comments":25,"created_at":"2026-10-29T13:50:39Z"},{"number":5031,"title":"[Engine] Qwen3.8 Flash-Next on two mixed NVIDIA cards (RTX 4090 24 GB + RTX 4080 16 GB)?","comments":19,"created_at":"2026-11-02T20:32:57Z"},{"number":4200,"title":"[Engine] Runtime control is unavailable inspect the private host log","comments":17,"created_at":"2026-10-26T08:40:11Z"},{"number":4350,"title":"[Engine] dsv41 on gfx1100 (HIP, self-built): switching back to a conversation after another one ran fails with \"illegal memory access\"","comments":16,"created_at":"2026-10-27T19:48:03Z"}],"discussions":[],"latest_release":{"name":"Strata 0.2.12","tag":"v0.2.12","published_at":"2026-11-07T08:54:40Z","body":"# Strata 0.2.12: a tuned DFlash2 drafter, faster prompts on 12 GB cards, Strix Halo and Mac, and tool-call fixes\n\nThe app and the Terminal are 0.2.12. Control and every engine package are 0.1.58. The update feed is sequence 122.\nThe DGX Spark (linux-aarch64) gets 0.2.12 too.\n\n## Highlights\n\n- **12 GB cards (RTX 3060 class):** UD-IQ4_XS-v31 prompts read about 3 times faster, 87 to 264 tokens per second at 2K.\n- **Strix Halo (Windows):** prompts x1.24 to x2.14 on UD-IQ4_XS-v31 and UD-Q4_K_XL.\n- **Mac (M5 Max):** output +10.3% to +12.7% and prompts +11.7% to +13.9% on UD-Q4_K_XL and UD-IQ4_XS-v31.\n- **Tuned DFlash2 drafter:** Strata's own copy of the helper model that guesses the next words; decode +2.4% to +5.0% on the RTX PRO 5000 and the DGX Spark.\n- **Radeon R9700:** 32K prompts +8.6% to +11.6%.\n- **Tool calls:** Claude Desktop's model name is accepted (#5312), plus fixes for #5409, #5276, #5281, #4622 and PRs #4828 and #4832.\n\n## Speed against 0.2.11\n\nTokens per second, 0.2.11 to 0.2.12. \"Output\" is the geometric mean over the kinds we tried where we have one; the worst kind is shown when it is a loss.\n\n| Machine, model | Prompt | Output |\n|---|---|---|\n| Mac M5 Max, UD-Q4_K_XL | 4K 1,406 to 1,605; 32K 1,350 to 1,524 | +12.7% (every kind up) |\n| Mac M5 Max, UD-IQ4_XS-v31 | 4K 1,440 to 1,649; 32K 1,373 to 1,543 | +10.3% (agent -0.9%) |\n| Strix Halo, UD-IQ4_XS-v31 | 4K 690 to 1,220; 32K 1,107 to 1,377 | code +10%, chat +3%; **tool call -9.5%** |\n| Strix Halo, UD-Q4_K_XL | 4K 597 to 1,267; 32K 1,034 to 1,370 | code +24%, chat +6%; **tool call -5.8%**, agent -2.5% |\n| RTX 3060 12 GB, UD-IQ4_XS-v31 | 2K 87 to 264; 32K 361 to 981 | code +37%, agent +34%; **JSON -18.9%** |\n| RTX PRO 5000, IQ3_S | 8K level (x1.002) | +2.4%; agent -1.5%, chat -1.0% |\n| DGX Spark, IQ3_S | see Known problems | +3.1%; chat -3.3% |\n| DGX Spark, UD-IQ4_XS-v31 | see Known problems | +5.0%; chat -0.4% |\n| Radeon R9700, IQ3_S (Auto now MTP) | 32K 3,076 to 3,304 | **code -10.5%, tool call -9.3%**, JSON -2.3% |\n| Radeon R9700, UD-IQ4_XS-v31 (Auto now MTP) | 32K 2,817 to 3,015 | **code -8.5%**, JSON -4.8%, tool call -3.2% |\n| 2 x RTX A6000, IQ3_S | level | split engine -0.40%, layer split +0.38% |\n\nHow it was measured: each machine ran 0.2.11 and the 0.2.12 candidate side by side, 2 to 3 rounds, a fresh engine start per run, the arms swapped in order each round, greedy answers of 300 tokens. Each version used the settings it writes for that machine, so the 0.2.12 column includes the new drafter choices. Speeds were not measured again on the final 0.1.58 packages; the engine source after the check changed only in the version number and two build fixes for old NVIDIA cards.\n\n## What changed for you\n\n- **Tuned DFlash2 drafter (0.34 GB, 0.52 GB on Intel Arc).** Auto now uses it on: Mac (UD-Q4_K_XL, UD-IQ4_XS-v31; DFlash2 is now the Mac default for these, confidence 0.7); NVIDIA cards with every expert on the card and the DGX Spark (IQ3_S, UD-IQ4_XS-v31); Strix Halo (IQ3_S, UD-Q4_K_XL, UD-IQ4_XS, UD-IQ4_XS-v31); Intel Arc A-series (IQ3_XXS); 12 GB NVIDIA cards (UD-IQ4_XS-v31). Back: `strata launch --drafting mtp`.\n- **Radeon gfx1201 (R9700, RX 9070): Auto starts MTP, not DFlash2,** because 11 of 50 DFlash2 engines on 0.2.11 stopped on the prompt. Line shown: \"DFlash2 is off by default on this graphics card for now; choose it in Drafting to use it.\" Back: `--drafting dflash2`.\n- **12 GB cards (Linux): the second start-up check no longer counts the engine's own memory as free,** so it stops switching to the slower file cache. When it still switches it says: `strata: expert reads now go through the file cache (X GiB of RAM free for them, Y GiB limit)`. On the RTX 3060, 0.2.11 switched in 3 of 3 starts and 0.2.12 in 0 of 3. Auto now starts MTP for IQ3_S and the tuned DFlash2 for UD-IQ4_XS-v31 on these cards. Back: `STRATA_FILE_TIER_HONEST=0`.\n- **A download with the wrong checksum is kept as `<name>.bad`, not deleted:** `<name> does not match the pinned SHA-256 (it is <"},"readme":true,"readme_zh":true,"news":[{"title":"Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s","url":"https://github.com/Niko1221/Strata","hn":"https://news.ycombinator.com/item?id=49953495","source":"Hacker News","lang":"en","date":"2026-10-04","points":940,"comments":420},{"title":"Strata: Qwen3.8-Flash-Next (125B Moe) on a 8GB+ Nvidia GPU","url":"https://github.com/Niko1221/Strata","hn":"https://news.ycombinator.com/item?id=49879253","source":"Hacker News","lang":"en","date":"2026-09-28","points":5,"comments":0},{"title":"Strata: Run a 125B-parameter AI model on a normal gaming PC","url":"https://github.com/Niko1221/Strata","hn":"https://news.ycombinator.com/item?id=49914465","source":"Hacker News","lang":"en","date":"2026-09-30","points":3,"comments":1},{"title":"Is Strata a Game Changer for Local LLMs? Considering Large MoE and Adding GPUs with the RTX 5090","url":"https://note.com/gyokuto15/n/nc0d4f19b53fc?hl=en","source":"note","lang":"en","date":"2026-11-05"},{"title":"【strata】strataのDFlash2はあまり速くならなかった話【RTX3060×2】","url":"https://note.com/catap_art3d/n/n37bac2c745fc","source":"note","lang":"zh","date":"2026-11-03"},{"title":"Open-Source Fix Slashes DeepSeek V4.1 Startup Time on Gaming GPUs","url":"https://startupfortune.com/open-source-fix-slashes-deepseek-v41-startup-time-on-gaming-gpus/","source":"Startup Fortune","lang":"en","date":"2026-11-01"},{"title":"普通人真能跑1-bit大模型了！Strata 0.2.9发布：AMD/NVIDIA/Intel全平台狂飙，32K长文本提速88%","url":"https://www.sohu.com/a/1095449677_122066679","source":"搜狐","lang":"zh","date":"2026-10-31"},{"title":"本地AI大模型Strata 0.2.9发布：处理速度提升88%，支持1-bit DeepSeek！","url":"https://www.sohu.com/a/1095449300_122004016","source":"搜狐","lang":"zh","date":"2026-10-31"},{"title":"I tried running Strata on an old custom-built PC with a 14-year-old CPU by plugging in an RTX 3060","url":"https://note.com/alpha2biker/n/n677582d71634?hl=en","source":"note","lang":"en","date":"2026-10-30"},{"title":"Intel Arc B580 runs 125B AI model nearly 3x faster with XeStrata in new test","url":"https://www.notebookcheck.net/Intel-Arc-B580-runs-125B-AI-model-nearly-3x-faster-with-XeStrata-in-new-test.1434389.0.html","source":"Notebookcheck","lang":"en","date":"2026-10-26"},{"title":"Strata：用显存+内存+硬盘协同调度，让普通电脑跑 1250 亿参数模型","url":"https://news.google.com/rss/articles/CBMiSEFVX3lxTE50dzhELWJBRXgxalVxRjdXQnd0c1I2MTNxZlFXS3NNY2hfNXFCcDRpQmZBRV9WXzNMcTlSNUFCYTZNRG0zTVJXMQ?oc=5","source":"黑海洋Wiki","lang":"zh","date":"2026-10-26"},{"title":"新浪网· - 手机新浪网","url":"https://news.google.com/rss/articles/CBMiY0FVX3lxTE5NN0IwcU0zX0hfeUJrTTY4blBoRlBkalVMeVVHcGthN25KVW9QM0M5SDhSWkgxcXlaS3QzcXlZZEt1YXRENFZNMnJDZjZ4YUJXeVdMMVo2ZDZ3LV95cFdDTnEtWQ?oc=5","source":"手机新浪网","lang":"zh","date":"2026-10-19"},{"title":"开发者Vibe两天产出Strata桌面管理端：本地大模型部署调参监控一站搞定","url":"https://news.google.com/rss/articles/CBMid0FVX3lxTE01MjczZDdpektQcGdLdU95OVduSC1saG1kTWc2ZGdOMWpvdHZCbGtmU2JKVlpGYmpNdFBQdHZGVVdSZGRHYUN0NTRIREhwSU5ZSElKdlVSbWkwbXMzWWFUR2YzVkwtOVlCLWViSHVRUHA4MENHVUJR?oc=5","source":"https://www.80aj.com/","lang":"zh","date":"2026-10-16"},{"title":"AI服务器需求推高存储成本：全球PC出货量下滑，DDR4平台获延长支持","url":"https://news.google.com/rss/articles/CBMiVEFVX3lxTFBnZXJWN3E5VDV3Njh1UHVvWmJfclJlZXVaU09QdGk5eGZNN2tQbWlsVnFrU2VuTW9TTHNPdTRwemY1RndSNkhoaFZ0UFlhMmQtYkFFVA?oc=5","source":"虎嗅","lang":"zh","date":"2026-10-10"},{"title":"Strata rewrites its GitHub history, 595 Claude co-author tags vanish","url":"https://news.google.com/rss/articles/CBMinAFBVV95cUxPWkNraFFtZTg5TkE5NmllMGVFQnFlaDJVemhvQ3RnZktEU2plRzJ0U1NfbUxTdExId09wWkVEWHdYUTFsdER0VkhoZS16ejk0cVJMMDhJRE9sNnBFdUluV2pLS0VJa202VXA0VUJWSEpmNXE4TV9UdV9STU9lWGhDS252bHYtdHRiLWp3MjVwakhtZHFRbnl4eDJlRjM?oc=5","source":"Pasquale Pillitteri","lang":"en","date":"2026-10-09"},{"title":"Strata 重写 GitHub 历史，595 条 Claude 署名消失，维护者确认重写","url":"https://news.google.com/rss/articles/CBMikAFBVV95cUxPWlB5YVhibklZcGRUaXRjYTV1VmQ5R2VKaXpzd200TnBXS0pvR3FBYl9EczF0UmdxV3FmRnk3Z3B2eE40OG1lektqaUtsMFNsMVM0TWs2aTliN01aY1hWSjdQRU5vUFBOeGVWOERVRC11eGxrV1RwSzJUaWxacGp6SFhPOXJvZlo3QmFpN0lLUVI?oc=5","source":"Pasquale Pillitteri","lang":"zh","date":"2026-10-09"},{"title":"12GB 显存显卡跑 125B Qwen3.8 模型：Strata 登场，单张 RTX 5070 跑出 94 词元 / 秒","url":"https://news.google.com/rss/articles/CBMiZEFVX3lxTFBZWUFQcG9uY09kSWNRUDgtYTJxWndGektjNHd4U3pTNk1JbXBUMlBxR2VGQ3dzUmc0UkNaUFNLMTdka3NQUm53ZkZpOHM0S0s1b1VrYXpiaGEwSUNNOE82dVN0QmU?oc=5","source":"17173","lang":"zh","date":"2026-10-06"},{"title":"Running a 125B-Parameter LLM on a Home PC: \"Strata\" Opens a New Frontier for Local AI","url":"https://news.google.com/rss/articles/CBMidkFVX3lxTE5KVDd3WFZNc2I0ZUpvWFNVLTNMQ0JQdW93TXdldGRVWF9iUmJPRkJZenh3eVBvN2dodDI0OG9WQWVTcnBaVzQ2QU9ncXhFYmRKR2MxTUxVWkM4YzJOeUtuTnVUWENoV2d4ZnE0M1NlQUYyZHdVN0E?oc=5","source":"BigGo Finance","lang":"en","date":"2026-10-05"},{"title":"Open-Source Tool Runs AI Locally","url":"https://news.google.com/rss/articles/CBMifEFVX3lxTE5yaHE0Y2N3WFRLQm9OYThqOVRRUl9jdkpuejFxbkkzdDZzZHN6YVYwR01BS0NQQ1RldDJ4RjVQb1dCb2VqNENSNUU3aktlaHZmZkFWLTVuSDVQSmlPOXRqUWM0YUV4WVZfcTg0akprM3BHM1ZVbXVtUmZIb1k?oc=5","source":"Open Source For You","lang":"en","date":"2026-10-05"},{"title":"'Strata' is an AI execution application that enables high-speed local execution of the quantized version of 125B's 'Qwen3.8-Flash-Next' on gaming PCs.","url":"https://news.google.com/rss/articles/CBMiakFVX3lxTE1tOGZzWjMxNHlVc3FVdEpJSlpUNEJJUTVnOXRTVU42ZWhmUWprWHNvZV9jMFd3X0l1bWNFNEdSQUE3NnF2NHZHM3pYRGpOWG5EdTNkSVJScV9kYkIxMlZLNWJESk16TWRQUWc?oc=5","source":"GIGAZINE","lang":"en","date":"2026-10-05"},{"title":"Strata lets a 125 billion parameter model run on a gaming GPU","url":"https://news.google.com/rss/articles/CBMikwFBVV95cUxNa2F0emhHU0c2Qkgxdmd2cVlHd0k5bG9rcVJCdFo2a3BPTkd4LXBGajNfRHZiNUtSRUtoeFpjeXJBNV9PcXB3NWswbmRmMmVMZmFNNXhVUE44cUJINTFDTUJBNGc5WGZXcS1oYUxWbXR4bVNEV2M5Q1Z1T1JnYk9nclRtVnllMEluT0hva3RGTm9rbWs?oc=5","source":"Startup Fortune","lang":"en","date":"2026-10-03"}],"stale":false,"snapshot_at":"2026-11-10T09:35:25Z","history_start":"2026-09-06"}