Models

This source gives a title and a link, nothing to quote.

They claim — 204 ms on one B200 (JEV-27B-VL 271 ms).

Published 2026-10-11

Fetched 2026-10-12 23:01 UTC · 3739ff1538c59c04

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Because text encoding only runs once per prompt, this saves 9–17 GB of VRAM with virtually zero impact on generation speed.

Published 2026-09-20

Fetched 2026-10-12 23:01 UTC · b9c64708f65f575b

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — [2026/07/21] 🤝 Thanks to the ms-swift community for their support, our model now supports training with ms-swift.

Published 2026-06-19

Fetched 2026-10-12 23:01 UTC · 62e6e76d595b1da6

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — As a chat model, the encoder's refusals fall from 88.9% to 1.2% in English and from 38% to 0% in Russian on held-out prompts.

Published 2026-10-09

Fetched 2026-10-10 23:01 UTC · 36c3a0b1565ee3bd

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — The text-to-image example uses 1680 × 2512 portrait resolution; the editing example uses 2048 × 2048.

Published 2026-10-09

Fetched 2026-10-09 23:01 UTC · 7cad2807a6dba4a4

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Qwen3.8-27B in under 8 GB, tuned to keep tool calling intact.

Published 2026-10-08

Fetched 2026-10-09 23:01 UTC · fd535ff0fdcb55b5

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — 55.0% fewer LLM calls, 21.0 ms per decision on an RTX 5090.

Published 2026-10-08

Fetched 2026-10-09 23:01 UTC · 53fb1f4f0b1a1150

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — top answer on 0.8% of text and 1.7% of audio rows.

Published 2026-10-05

Fetched 2026-10-11 23:01 UTC · c0297536843caeee

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Restoration tiles large outputs into overlapping 1024×1024 windows.

Published 2026-10-05

Fetched 2026-10-10 23:01 UTC · fa4451995d49a027

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — model, that boost the Decision Index score from 56.4 to 61.56, now outperforming Jev by more than 3.5 points, while using the same backbone.

Published 2026-10-05

Fetched 2026-10-10 23:01 UTC · 2db5bbc68088a7bf

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — For NInfer W4A4, the 2 LCB empty answers are problems 11 and 15, both truncated with no code; the 2 GPQA empty answers are problems 79 and 127, both truncated with an empty pred.

Published 2026-10-04

Fetched 2026-10-09 23:01 UTC · 1c8c0b72263dc3b9

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — OrcaSAQ2 performs strongly on long-horizon workloads relative to models in its deployment and parameter class, despite operating from a 12.3 GB checkpoint.

Published 2026-09-24

Fetched 2026-09-26 23:01 UTC · 82644ada59f66f78

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Typed accuracy: Choice 71.33% (428/600), Noul 80.67% (484/600), Score 68.88% (551/800).

Published 2026-09-23

Fetched 2026-09-27 23:01 UTC · f83624fe1b9581b4

Torn from — the model card

This source gives a title and a link, nothing to quote.

Published 2026-09-23

Fetched 2026-09-26 23:01 UTC · b267fd54205b20a2

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Terminal-Bench 2.1 (87.6%), 4–6× faster than Jev.

Published 2026-09-21

Fetched 2026-09-25 23:01 UTC · 96c8aa6a5329130e

Torn from — the model card

This source gives a title and a link, nothing to quote.

Published 2026-09-21

Fetched 2026-09-25 23:01 UTC · 436cd7dd55e01e86

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Supra2-IMG is a tiny 100M parameters text-to-image (T2I) model that has been trained from scratch on high-quality synthetic data and delivers state-of-the-art image quality for its size.

Published 2026-09-21

Fetched 2026-09-27 23:01 UTC · e006b2ea9024da28

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — MacBook Air with the Apple M2 (4 performance and 4 efficiency CPU cores, 10-core GPU), 16 GB unified memory, macOS 15.

Published 2026-09-18

Fetched 2026-09-26 23:01 UTC · 6144a78a085e7fc8

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — We evaluated NaviDC-OCR with its native weights, achieving better results than MinerU 2.5 Pro and PaddleOCR-VL 1.6.

Published 2026-08-14

Fetched 2026-09-28 05:05 UTC · e9a2dd67be526bb6

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — bf16 variants are half the size; head top-1 agreement with fp32 is 98.6%).

Published 2026-09-13

Fetched 2026-09-15 23:01 UTC · 1bd7034ce20966aa

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — (<1% loss) and as a result getting a x1.95 speed-up on several tasks.

Published 2026-09-11

Fetched 2026-09-15 23:01 UTC · fae0d09608b75058

Torn from — the model card

This source gives a title and a link, nothing to quote.

Published 2026-09-11

Fetched 2026-09-15 23:01 UTC · 65418212282ccac7

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — The refusal circuitry is surgically removed while every capability-critical component (routed experts, Engram memory, CSA2 sparse attention, DSpark draft head, vision tower, router gates,…

Published 2026-09-10

Fetched 2026-09-13 23:03 UTC · aba564a993aae2f3

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — TWIN TURBO: Closed source level of intelligence with an arc-c at 709 (this is instruct "medium", reasoning is even higher), while using smaller quants too.

Published 2026-09-10

Fetched 2026-09-14 23:01 UTC · 68058414275bc9a3

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — (<1% loss) and as a result getting a x1.95 speed-up on several tasks.

Published 2026-09-08

Fetched 2026-09-14 23:01 UTC · b4d8039f21680101

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — Within the 10B-scale general-purpose VLMs compared in this release blog, ZDTaichu5.0-9B retains first-tier general visual understanding while supporting spatial reasoning, high-level embo…

Published 2026-09-04

Fetched 2026-09-15 23:01 UTC · 5179574dcefd9e42

Torn from — the model card

This source gives a title and a link, nothing to quote.

Published 2024-07-18

Fetched 2026-09-14 23:01 UTC · bad7065e70b23d0e

Torn from — the listing only (the original would not open)

This source gives a title and a link, nothing to quote.

They claim — It is a dense 2B Transformer that scales up the same training recipe, built for on-device, local deployment, and resource-constrained scenarios, reaching 2B-class open-source SOTA.

Published 2026-09-06

Fetched 2026-09-07 23:01 UTC · 80cbd6d3b1d56d8d

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — 🛡️ Full Base Capability Retention: 100% preserves MiniMax-H3's original prompt adherence, style adaptability, and base multimodal generation strength.

Published 2026-09-05

Fetched 2026-09-06 23:01 UTC · 740cf7d9fea002ec

Torn from — the model card

This source gives a title and a link, nothing to quote.

They claim — It is a dense 2B Transformer that scales up the same training recipe, built for on-device, local deployment, and resource-constrained scenarios, reaching 2B-class open-source SOTA.

Published 2026-09-05

Fetched 2026-09-08 16:21 UTC · 8d445192f84dfbcc

Torn from — the model card