bytedance-research/Lance
Any-to-Any • Updated • 438 • 475
Alex Hant
hardhant
AI & ML interests
None yet
Recent Activity
liked a model 1 day ago
bytedance-research/Lance liked a Space 2 days ago
TencentARC/Pixal3D liked a Space 5 days ago
Supertone/supertonic-3Organizations
None yet
liked a
model 1 day ago
liked a
Space 2 days ago
liked a
Space 5 days ago
liked a
model 7 days ago
reacted to danielhanchen's post with 🔥 7 days ago
Post
5644
We’re excited to announce that Unsloth has joined the PyTorch Ecosystem! 🔥🦥
Unsloth is an open-source project that makes training & running models more accurate and faster with less compute. Our mission is to make local AI accessible to everyone. Thanks to all of you for making this possible! 💕
Blog: https://unsloth.ai/blog/pytorch
GitHub: https://github.com/unslothai/unsloth
Unsloth is an open-source project that makes training & running models more accurate and faster with less compute. Our mission is to make local AI accessible to everyone. Thanks to all of you for making this possible! 💕
Blog: https://unsloth.ai/blog/pytorch
GitHub: https://github.com/unslothai/unsloth
- 2 replies
liked a
model 8 days ago
liked a
model 10 days ago
reacted to DedeProGames's post with 🔥 12 days ago
Post
5048
🚀 Introducing the GRM-2.6 Family
The GRM-2.6 family is a new generation of reasoning-focused models from Orion LLM Labs, built for difficult tasks, coding, STEM, terminal agents, and advanced local AI workflows.
GRM-2.6-Plus is the main high-capability model in the family: a 27B-class reasoning model based on Qwen3.6, designed for strong structured reasoning, coding, agentic use, and practical local deployment.
GRM-2.6-Opus builds on GRM-2.6-Plus as a merge with an Opus-style reasoning distilled model, improving structured reasoning behavior, terminal-agent workflows, coding ability, and complex problem solving.
Both models are designed for users who want powerful reasoning models that remain practical for research, local inference, coding, and agent experiments.
Models:
GRM-2.6-Plus: OrionLLM/GRM-2.6-Plus
GRM-2.6-Opus: OrionLLM/GRM-2.6-Opus
Organization:
OrionLLM
The GRM-2.6 family is a new generation of reasoning-focused models from Orion LLM Labs, built for difficult tasks, coding, STEM, terminal agents, and advanced local AI workflows.
GRM-2.6-Plus is the main high-capability model in the family: a 27B-class reasoning model based on Qwen3.6, designed for strong structured reasoning, coding, agentic use, and practical local deployment.
GRM-2.6-Opus builds on GRM-2.6-Plus as a merge with an Opus-style reasoning distilled model, improving structured reasoning behavior, terminal-agent workflows, coding ability, and complex problem solving.
Both models are designed for users who want powerful reasoning models that remain practical for research, local inference, coding, and agent experiments.
Models:
GRM-2.6-Plus: OrionLLM/GRM-2.6-Plus
GRM-2.6-Opus: OrionLLM/GRM-2.6-Opus
Organization:
- 2 replies
reacted to danielhanchen's post with 🔥 15 days ago
Post
8766
We made a guide on how to run open LLMs in Claude Code, Codex and OpenClaw.
Use Gemma 4 and Qwen3.6 GGUFs for local agentic coding on 24GB RAM
Run with self-healing tool calls, code execution, web search via the Unsloth API endpoint and llama.cpp
Guide: https://unsloth.ai/docs/basics/api
Use Gemma 4 and Qwen3.6 GGUFs for local agentic coding on 24GB RAM
Run with self-healing tool calls, code execution, web search via the Unsloth API endpoint and llama.cpp
Guide: https://unsloth.ai/docs/basics/api
reacted to DedeProGames's post with 🔥 15 days ago
Post
8569
GRaPE 2 Pro is now available.
SL-AI/GRaPE-2-Pro
This is the flagship model of the GRaPE 2 family and the largest model I have trained to date, sitting at 27B parameters. It is built on Qwen3.5-27B and trained on a closed-source proprietary dataset, with roughly half of post-training focused on code and the rest split between STEAM subjects and structured logical reasoning. It punches seriously above its weight class.
GRaPE 2 Pro supports multimodal input (image + text) and features 6 thinking modes via the
It also runs on consumer hardware. You can get it going with as low as 12GB of VRAM on a quantized build.
If you want to try it out and give feedback, that would be really appreciated. Email us at
SL-AI/GRaPE-2-Pro
This is the flagship model of the GRaPE 2 family and the largest model I have trained to date, sitting at 27B parameters. It is built on Qwen3.5-27B and trained on a closed-source proprietary dataset, with roughly half of post-training focused on code and the rest split between STEAM subjects and structured logical reasoning. It punches seriously above its weight class.
GRaPE 2 Pro supports multimodal input (image + text) and features 6 thinking modes via the
minimal, all the way up to xtra-Hi for deep, extended thought on hard problems. For most agentic use, auto or low is the move to keep things snappy.It also runs on consumer hardware. You can get it going with as low as 12GB of VRAM on a quantized build.
If you want to try it out and give feedback, that would be really appreciated. Email us at
contact@skinnertopia.com 
- 1 reply
reacted to ManniX-ITA's post with 🚀 19 days ago
Post
3012
🚀 Two releases this week pushing merge methodology forward.
▶ Qwen3.6-27B-Omnimerge-v4-MLP
ManniX-ITA/Qwen3.6-27B-Omnimerge-v4
Same-base DARE-TIES merge of Qwen3.6-27B + 3 fine-tunes (rico03 Claude distill, Esper3.1, kai-os Opus reasoning anchor) via my Omnimerge_v2 method (OBIM-lite + DAREx-q + EMR election).
Hit a Qwen3.6-specific fragility: hyperparams that work flawlessly on 3.5 produced 80% unclosed-<think> on 3.6, collapsing pass@1 to ~20%. Per-tensor delta forensics localized the failure to mlp.{gate,up,down}_proj in
layers 27–52. Fix: MLP-passthrough surgery — copy MLPs verbatim from base, keep merged attn + linear_attn. Leak → 0%.
Q6_K results (vs Qwen3.6 base / vs Omnimerge-v2 on Qwen3.5):
• HumanEval: 84.76% (= base, +5.49 pp vs v2)
• MBPP corrected: 73.40% (+15.80 pp vs base, ≈ v2)
• GPQA Diamond: ~84.75% partial 192/198 (+15.5 pp vs v2)
▶ Qwen3.5-4B Importance-Signal Study (M1..M5)
Controlled 5-way comparison: same Qwen3.5-4B base, same 2 fine-tunes (Jackrong Claude-4.5 distill + Crow Opus-4.6 distill), only the importance signal driving DARE-TIES sparsification varies.
Q6_K HE / MBPP pass@1:
• M1 Vanilla DARE-TIES → 51.22 / 47.00
• M2 OMv2 (no signal) → 52.44 / 49.40
• M3 OMv2 + Fisher → 57.93 🥇 / 48.80
• M4 mergekit ex-LRP (PR #682) → 51.22 / 49.40
• M5 OMv2 + LRP → 53.05 / 51.40 🥇
Findings: Fisher wins HE (+4.88 pp over vanilla), LRP wins MBPP (+2.60 pp). Both signals + Omnimerge_v2 recipe beat vanilla. To make multimodal-LM ex-LRP work end-to-end against Qwen3_5ForConditionalGeneration, I filed
5 patches against arcee-ai/mergekit PR #682 + 1 against rachtibat/lxt.
All five Mx checkpoints + Fisher/LRP signal safetensors + reproducer scripts published.
▶ Qwen3.6-27B-Omnimerge-v4-MLP
ManniX-ITA/Qwen3.6-27B-Omnimerge-v4
Same-base DARE-TIES merge of Qwen3.6-27B + 3 fine-tunes (rico03 Claude distill, Esper3.1, kai-os Opus reasoning anchor) via my Omnimerge_v2 method (OBIM-lite + DAREx-q + EMR election).
Hit a Qwen3.6-specific fragility: hyperparams that work flawlessly on 3.5 produced 80% unclosed-<think> on 3.6, collapsing pass@1 to ~20%. Per-tensor delta forensics localized the failure to mlp.{gate,up,down}_proj in
layers 27–52. Fix: MLP-passthrough surgery — copy MLPs verbatim from base, keep merged attn + linear_attn. Leak → 0%.
Q6_K results (vs Qwen3.6 base / vs Omnimerge-v2 on Qwen3.5):
• HumanEval: 84.76% (= base, +5.49 pp vs v2)
• MBPP corrected: 73.40% (+15.80 pp vs base, ≈ v2)
• GPQA Diamond: ~84.75% partial 192/198 (+15.5 pp vs v2)
▶ Qwen3.5-4B Importance-Signal Study (M1..M5)
Controlled 5-way comparison: same Qwen3.5-4B base, same 2 fine-tunes (Jackrong Claude-4.5 distill + Crow Opus-4.6 distill), only the importance signal driving DARE-TIES sparsification varies.
Q6_K HE / MBPP pass@1:
• M1 Vanilla DARE-TIES → 51.22 / 47.00
• M2 OMv2 (no signal) → 52.44 / 49.40
• M3 OMv2 + Fisher → 57.93 🥇 / 48.80
• M4 mergekit ex-LRP (PR #682) → 51.22 / 49.40
• M5 OMv2 + LRP → 53.05 / 51.40 🥇
Findings: Fisher wins HE (+4.88 pp over vanilla), LRP wins MBPP (+2.60 pp). Both signals + Omnimerge_v2 recipe beat vanilla. To make multimodal-LM ex-LRP work end-to-end against Qwen3_5ForConditionalGeneration, I filed
5 patches against arcee-ai/mergekit PR #682 + 1 against rachtibat/lxt.
All five Mx checkpoints + Fisher/LRP signal safetensors + reproducer scripts published.
- 1 reply
reacted to danielhanchen's post with 🤗 21 days ago
reacted to ajibawa-2023's post with 🚀 about 1 month ago
Post
6109
Go-Code-Large
Dataset: ajibawa-2023/Go-Code-Large
Go-Code-Large is a large-scale corpus of Go (Golang) programming language source code, comprising 316,427 code samples stored in .jsonl format. The dataset is designed to support research and development in large language model (LLM) pretraining, static analysis, cloud-native systems, and modern backend software engineering.
By offering a focused and curated dataset for Go, this corpus enables experimentation in concurrent programming, distributed systems, and performance-oriented backend services—domains where Go is widely adopted.
Go-Code-Large addresses the relative scarcity of large, language-specific datasets for Go, enabling targeted research into idiomatic Go patterns, concurrency primitives, and scalable system design.
Dataset: ajibawa-2023/Go-Code-Large
Go-Code-Large is a large-scale corpus of Go (Golang) programming language source code, comprising 316,427 code samples stored in .jsonl format. The dataset is designed to support research and development in large language model (LLM) pretraining, static analysis, cloud-native systems, and modern backend software engineering.
By offering a focused and curated dataset for Go, this corpus enables experimentation in concurrent programming, distributed systems, and performance-oriented backend services—domains where Go is widely adopted.
Go-Code-Large addresses the relative scarcity of large, language-specific datasets for Go, enabling targeted research into idiomatic Go patterns, concurrency primitives, and scalable system design.
- 2 replies
reacted to DedeProGames's post with 🚀 about 1 month ago
Post
3483
🔥 GRM-2.5 - The most POWERFUL model for local inference
The GRM-2.5 is the newest model from Orion LLM Labs. It has consistent RAW reasoning and is capable of generating very precise responses, similar to large models, while maintaining a parameter size of 4b.
The GRM-2.5 family consists of these models:
OrionLLM/GRM-2.5 (4b)
OrionLLM/GRM-2.5-Air (0.8b)
Furthermore, the GRM-2.5 is the best option for local agentic environments, being very good in code, terminal agent, etc. It is capable of generating 1000 lines of consistent code and programming like large models.
The GRM-2.5 is the best base for FineTune to date and has vision, which means it can interpret images and videos.
The GRM-2.5 is the newest model from Orion LLM Labs. It has consistent RAW reasoning and is capable of generating very precise responses, similar to large models, while maintaining a parameter size of 4b.
The GRM-2.5 family consists of these models:
OrionLLM/GRM-2.5 (4b)
OrionLLM/GRM-2.5-Air (0.8b)
Furthermore, the GRM-2.5 is the best option for local agentic environments, being very good in code, terminal agent, etc. It is capable of generating 1000 lines of consistent code and programming like large models.
The GRM-2.5 is the best base for FineTune to date and has vision, which means it can interpret images and videos.
- 1 reply
reacted to danielhanchen's post with 🔥 about 1 month ago
Post
2835
Qwen3.6-35B-A3B can now be run locally! 💜
The model is the strongest mid-sized LLM on nearly all benchmarks.
Run on 23GB RAM via Unsloth Dynamic GGUFs.
GGUFs to run: unsloth/Qwen3.6-35B-A3B-GGUF
Guide: https://unsloth.ai/docs/models/qwen3.6
The model is the strongest mid-sized LLM on nearly all benchmarks.
Run on 23GB RAM via Unsloth Dynamic GGUFs.
GGUFs to run: unsloth/Qwen3.6-35B-A3B-GGUF
Guide: https://unsloth.ai/docs/models/qwen3.6
- 13 replies
reacted to imnotkitty's post with 🔥 about 1 month ago
Post
2611
Just tried tencent/HY-World-2.0 — a multimodal world model that takes in text or a single image and generates editable 3D scenes.
Unlike Google's Genie and HY-World 1.5, v2.0 generates engine-ready 3D content:
🎮 Direct import into Unreal Engine and Unity — no format wrangling
🧊 Supports multiple 3D asset formats: Mesh, 3DGS, point cloud, etc.
✏️ Fully editable — not a baked video, but actual geometry you can modify
🤖 Also usable for embodied simulation environments
Basically: from "AI generates a world you can look at" → "AI generates a world you can ship."
Unlike Google's Genie and HY-World 1.5, v2.0 generates engine-ready 3D content:
🎮 Direct import into Unreal Engine and Unity — no format wrangling
🧊 Supports multiple 3D asset formats: Mesh, 3DGS, point cloud, etc.
✏️ Fully editable — not a baked video, but actual geometry you can modify
🤖 Also usable for embodied simulation environments
Basically: from "AI generates a world you can look at" → "AI generates a world you can ship."
reacted to Benedictat's post with 🚀 about 1 month ago
Post
2749
Hunyuan HY-World 2.0 Open-Sourced | Unified SOTA for 3D Generation / Reconstruction / Simulation
HY-World 2.0 is a unified 3D world model supporting multimodal inputs including text and images.
Its end-to-end framework simultaneously performs 3D understanding, scene generation, and geometric reconstruction.
Based on HY-Pano-2.0, the model enables panorama generation without camera parameters
It ensures geometric consistency via spatial agents and trajectory planning, and achieves joint 3DGS & Mesh representation with WorldMirror 2.0, reaching SOTA performance in novel view synthesis and 3D reconstruction
Unlike Genie 3 and HY-World 1.5, which only output videos, HY-World 2.0 directly generates editable 3D assets, better meeting real-world research and simulation demands
HY-World 2.0 is a unified 3D world model supporting multimodal inputs including text and images.
Its end-to-end framework simultaneously performs 3D understanding, scene generation, and geometric reconstruction.
Based on HY-Pano-2.0, the model enables panorama generation without camera parameters
It ensures geometric consistency via spatial agents and trajectory planning, and achieves joint 3DGS & Mesh representation with WorldMirror 2.0, reaching SOTA performance in novel view synthesis and 3D reconstruction
Unlike Genie 3 and HY-World 1.5, which only output videos, HY-World 2.0 directly generates editable 3D assets, better meeting real-world research and simulation demands
- 1 reply
reacted to prithivMLmods's post with ❤️ about 1 month ago
Post
6217
A new comparator on Spaces showcases Standard FLUX.2 Decoder vs. FLUX.2 Small Decoder. The Small Decoder is ~1.4× faster, uses ~1.4× less VRAM, and maintains near-identical image quality. It has ~28M parameters with narrower channels [96, 192, 384, 384] vs. [128, 256, 512, 512], and the demo supports sequence generation by running both decoders simultaneously and comparing the results side by side.
🤗 Comparator: https://huggingface.co/spaces/prithivMLmods/Flux.2-4B-Decoder-Comparator
🔗 FLUX.2-small-decoder: black-forest-labs/FLUX.2-small-decoder
🔗 GitHub: https://github.com/PRITHIVSAKTHIUR/Flux.2-4B-Encoder-Comparator
🚁 Collection: https://huggingface.co/collections/prithivMLmods/image-generation-apps-collection
🤗 > App built on the Gradio SDK. To learn more, visit the app page or the respective model pages.
🤗 Comparator: https://huggingface.co/spaces/prithivMLmods/Flux.2-4B-Decoder-Comparator
🔗 FLUX.2-small-decoder: black-forest-labs/FLUX.2-small-decoder
🔗 GitHub: https://github.com/PRITHIVSAKTHIUR/Flux.2-4B-Encoder-Comparator
🚁 Collection: https://huggingface.co/collections/prithivMLmods/image-generation-apps-collection
🤗 > App built on the Gradio SDK. To learn more, visit the app page or the respective model pages.
reacted to omarkamali's post with 👍 about 1 month ago
Post
4578
We got Qwen 3.5 to count Rs in Strawberry correctly! 🚨
Building on Sawtone, we’ve been testing a different way to feed language into an LLM to build the next generation of multilingual AI.
The usual setup gives the model tokenized text and asks it to perform various linguistic tasks. That works surprisingly well, until it doesn’t. Accents disappear. Words get mangled. Internal structure gets blurred away. And the cost of that gets higher once you move into multilingual and lower-resource settings.
So we tried adding a second path.
In addition to the normal text input, the model also receives Sawtone: a byte-level word representation that preserves how a word is written, how it sounds, and how it is structured.
Same LLM. Better interface.
In this proof of concept with Qwen 3.5 0.8B, that pushed our eval from 64% to 88%. The gains showed up exactly where tokenized models usually get shaky: diacritics, character order, exact spelling, and other form-sensitive behavior.
Sawtone itself is tokenizer-free, byte-level, and pre-trained across 507 languages.
Still early, but promising!
Building on Sawtone, we’ve been testing a different way to feed language into an LLM to build the next generation of multilingual AI.
The usual setup gives the model tokenized text and asks it to perform various linguistic tasks. That works surprisingly well, until it doesn’t. Accents disappear. Words get mangled. Internal structure gets blurred away. And the cost of that gets higher once you move into multilingual and lower-resource settings.
So we tried adding a second path.
In addition to the normal text input, the model also receives Sawtone: a byte-level word representation that preserves how a word is written, how it sounds, and how it is structured.
Same LLM. Better interface.
In this proof of concept with Qwen 3.5 0.8B, that pushed our eval from 64% to 88%. The gains showed up exactly where tokenized models usually get shaky: diacritics, character order, exact spelling, and other form-sensitive behavior.
Sawtone itself is tokenizer-free, byte-level, and pre-trained across 507 languages.
Still early, but promising!
- 5 replies
reacted to SeaWolf-AI's post with 👀 about 1 month ago
Post
3043
Why This Matters — David Defeats Goliath
MODEL: FINAL-Bench/Darwin-4B-David
SPACE: FINAL-Bench/Darwin-4B-david
We're releasing Darwin-4B-David, the first second-generation model in the Darwin Opus family. By evolving an already-evolved model, it achieves 85.0% on GPQA Diamond — surpassing its 58.6% original ancestor and even gemma-4-31B (84.3%) — with just 4.5B parameters.
Second-Generation Evolution
Most merges start from a base model and produce a single offspring. Darwin-4B-David breaks this pattern. The Father (Darwin-4B-Opus) was already evolved from gemma-4-E4B-it with Claude Opus reasoning distillation — a Gen-1 model. The Mother (DavidAU's DECKARD-Expresso-Universe) brings Unsloth deep tuning across 5 in-house datasets with thinking mode by default. Crossbreeding these two produced the first Gen-2 Darwin model.
Darwin V6's Model MRI scanned both parents across all 42 layers, assigning independent optimal ratios per layer. The Mother's creativity and Korean language hotspot (Layer 22-25, weight 0.95) was maximally absorbed, while the Father's reasoning core (Layer 30-40, weight 0.48) was preserved. This is "Merge = Evolve" applied recursively — evolution of evolution.
Benchmarks
Darwin-4B-David scores 85.0% on GPQA Diamond (+26.4%p over original 58.6%), evaluated generatively with maj@8 (8 generations per question, majority vote), Epoch AI prompt format, thinking mode enabled, 50 sampled questions. On ARC-Challenge (25-shot, loglikelihood), both score 64.93% — expected, as loglikelihood doesn't capture thinking-mode reasoning differences.
Why This Matters
gemma-4-31B (30.7B) scores 84.3%. Darwin-4B-David surpasses it at 1/7th the size — no training, no RL, just 45 minutes of MRI-guided DARE-TIES on one H100. The name "David" honors Mother creator DavidAU and evokes David vs. Goliath.
MODEL: FINAL-Bench/Darwin-4B-David
SPACE: FINAL-Bench/Darwin-4B-david
We're releasing Darwin-4B-David, the first second-generation model in the Darwin Opus family. By evolving an already-evolved model, it achieves 85.0% on GPQA Diamond — surpassing its 58.6% original ancestor and even gemma-4-31B (84.3%) — with just 4.5B parameters.
Second-Generation Evolution
Most merges start from a base model and produce a single offspring. Darwin-4B-David breaks this pattern. The Father (Darwin-4B-Opus) was already evolved from gemma-4-E4B-it with Claude Opus reasoning distillation — a Gen-1 model. The Mother (DavidAU's DECKARD-Expresso-Universe) brings Unsloth deep tuning across 5 in-house datasets with thinking mode by default. Crossbreeding these two produced the first Gen-2 Darwin model.
Darwin V6's Model MRI scanned both parents across all 42 layers, assigning independent optimal ratios per layer. The Mother's creativity and Korean language hotspot (Layer 22-25, weight 0.95) was maximally absorbed, while the Father's reasoning core (Layer 30-40, weight 0.48) was preserved. This is "Merge = Evolve" applied recursively — evolution of evolution.
Benchmarks
Darwin-4B-David scores 85.0% on GPQA Diamond (+26.4%p over original 58.6%), evaluated generatively with maj@8 (8 generations per question, majority vote), Epoch AI prompt format, thinking mode enabled, 50 sampled questions. On ARC-Challenge (25-shot, loglikelihood), both score 64.93% — expected, as loglikelihood doesn't capture thinking-mode reasoning differences.
Why This Matters
gemma-4-31B (30.7B) scores 84.3%. Darwin-4B-David surpasses it at 1/7th the size — no training, no RL, just 45 minutes of MRI-guided DARE-TIES on one H100. The name "David" honors Mother creator DavidAU and evokes David vs. Goliath.