OpenMOSE (@_m0se_) header
OpenMOSE (@_m0se_) profile picture

Rice Farmer, Fish Farmer, Welder(TIG,MIG), Electrician, Programming, Rocm RWKV is all you need. 元高専生です https://teapawt.net/t.co/XICxvc5uor https://teapawt.net/t.co/NShkwbjoVS

Joined February 2023
Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)Photo by OpenMOSE (@_m0se_)
Replying to @_m0se_
please use with your codex.😀 3bit model. minimum system configuration is 1 x Radeon W7900 + 128GB of RAM. github.com/OpenMOSE/Strata-G…
3
3
16
1,825
Update Strata-GLM-AMD 3bpw 2x AMD w7900(48GB) + 128GB RAM Pre-fill 1200tps+, Decode 40-50tps now improved cache hit rate w/ Indexer-assisted Block Window Sparse Attention
heeeeey!!! Strata + GLM 5.3 Flash 320B running fully local on 2x AMD Radeon W7900. 1000+ tps pre-fill(w/ parallel pre-fill) 30+ tps decode (w/ MTP), 256k ctx, tool calls working. coming indexer-free hybrid sparse attention mode Local LLMs just got serious🚀
1
4
376
heeeeey!!! Strata + GLM 5.3 Flash 320B running fully local on 2x AMD Radeon W7900. 1000+ tps pre-fill(w/ parallel pre-fill) 30+ tps decode (w/ MTP), 256k ctx, tool calls working. coming indexer-free hybrid sparse attention mode Local LLMs just got serious🚀
15
31
4
269
29,652
I think it would run on NVIDIA hardware too if MMQ were ported, but sadly, I don't have a Blackwell 96GB...
4
1,162
Will commit soon
1,492
Strata + GLM-5.3 Flash😀 still 20tps decode...
2
281
Strataやっと触る時間できた IQ3_XXS 2 x 7800XT 32GB + 128GB 67.5tps (たぶん64GBいける) 2 x W7900 96GB + 128GB 100tps A6000+3090 72GB +128GB 112tps 7800XTの中古実売が 6万くらいだから、 結構コスパ良い説
6
528
論文デスマーチ
1
188
最近は、いろいろあってクローズド研究ばかりですが、、 オープン研究も加速したいと思っています。 線形の旅は続きます。
5
178
JevとModernBertの違いについて教えて
1
5
350
OpenMOSE retweeted
1/ How can we make Low-Rank Matrices behave more like Full-Rank Matrices—and achieve higher rank efficiency? We release our new paper: NoRA: Normalized Low-Rank Adaptation. Paper: arxiv.org/abs/2608.31036
4
24
6
143
23,052
OpenMOSE retweeted
RWKV-7 G1j (100% RNN) release 🙂 much better at agent/coding/STEM and everything. Demo: rwkv.com Weights: huggingface.co/BlinkDL/rwkv7…
RWKV-7 G1i (100% RNN) release 🙂 Demo: rwkv.com Weights: huggingface.co/BlinkDL/rwkv7…
3
12
7
128
25,380
GLM-5.3-Flash-REAP-250B-A18B REAPモデルリリースしました。 Opencode + 日英中テキストの多段階評価合成プルーニングをしています。 MoEの冗長性が低いため、20%にとどまりましたが、 OpenCodeはそれなりに動作します。 huggingface.co/OpenMOSE/GLM-…
5
1
37
6,030
GLM5.3 Flash REAP 28%ダイエット版 英語、日本語 + OpenCode特化キャリブレーション MoEの分散が強すぎて、汎用プルーニングはあきらめました mi325x TP=4で、 200tpsくらい
10
783
GLM5.3 Flash のMoEは分散の極みで、 とっても刈り取るの難しい
4
253
戦争に使われる弾薬をRTX 5090換算w
打仗有多烧钱? 换成RTX 5090显卡你就知道了
5
109
6
362
29,096
GLM 5.3 Flash 魔改造します
1
10
332
またアーキテクチャ変えるらしい
The next-gen architecture powering Qwen4 is now here! ✨ Get ready for the open release of Qwen3.8-Flash-Next 🚀 The countdown starts now! ⏳🔥modelscope.cn/models/Qwen/Qw…
3
207