🚀 ZDTaichu5.0-9B is now on ModelScope! 🤖 An on-device multimodal model from TaichuAI. At 9B parameters it runs on a single GPU and brings spatial reasoning, embodied AI and agentic tool use to edge deployment. Qwen3.5-9B backbone + C-RADIOv4-H vision encoder, 128K context, any-resolution image and video input. 🧭 Spatial reasoning: leads the compared 10B-scale open VLMs (Qwen3.5-9B, STEP3-VL-10B, gemma4-8B-E4B) and scores above Gemini 3 Pro, Grok 4 and GPT-5.2 on ViewSpatial, MMSI-Bench and MindCube-tiny 🛠️ Agent: highest among the compared open models on TAU2-Bench, Claw-Eval and IFEval 📄 First-tier results on documents, charts, OCR, visual math and video, with a ready-to-use vLLM branch and Docker image 🧠 Entropy-Gated Adaptive Recurrent Reasoning: extra latent refinement steps go only to the hard tokens

Okay, so specialized decision models are becoming the rage. Bespoke Labs just released Nimble, an open 9B decision model built from Qwen3.5-9B. And apparently they built it in ONE DAY. The recipe: 🧠 Qwen3.5-9B 🎯 LoRA fine-tune 📚 only 2,676 training examples ❌ no Jev distillation ❌ no RL ✅ open model ✅ open data ✅ open training recipe Instead of generating explanations or JSON, Nimble scores the allowed answers: YES / NO → A / B / C → route 1 / 2 / 3 → severity 1–5 Then returns the choice + probabilities. Example of Bespoke's 324-example held-out test: Qwen3.5-9B → 66.36% Qwen3.8-27B → 84.88% 🔥 Nimble-9B → 90.12% Jev → 93.21% And because it isn't generating a bunch of tokens: ⚡ ~106 ms median on H100 🍎 ~444 ms median on an M5 Pro 64GB 👈 👀 🎯 So this is the key, giant reasoning model doesn't need to answer every tiny agent decision. Let the big model think, then let a small local model handle the ... 🛠️ tool selection 🧭 routing ✅ verification 🚨 policy decisions 📊 scoring 🔄 retry / stop Fast. Local. No API call. ⚠️ Bespoke says this is a narrow synthetic 324-example evaluation, NOT a standardized System One benchmark. 🔗 GH: /bespokelabsai/nimble

💥 Okay, yesterday we had MiniCPM5-2B, but today @OpenSquilla turned Qwen3.5-9B into a ~5.7GB LOCAL agent model, and their training method is interesting. The new NeoHorse-1-9B family is a post-trained Qwen3.5-9B specifically for agents, tool use + coding. It learned from actual multi-model agent execution traces. Setup ... 🔧 tool calls ❌ failed attempts 🔀 model handoffs 🧠 changed plans ✅ environment-verified successes Then they fed those experiences back into the 9B model. Their reported results vs Qwen3.5-9B: 🤖 QwenClaw: 44.04 → 48.73 📌 PinchBench: 74.55 → 82.25 🔄 VitaBench: 31.25 → 42.25 🛠 BFCL v4: 64.88 → 67.43 💻 HumanEval: 92.68 → 98.17 Overall: 65.60 → 69.04 This is the model I'm most interested in for ThumbLLM 🧠 ~9B parameters 💾 Q4_K_M GGUF ~5.7GB 🦙 llama.cpp 🟢 Ollama 🖥️ LM Studio 🤖 Hermes / OpenClaw 📚 262K native context ⚖️ Apache 2.0 So you can run this entirely locally on CPU or GPU. This isn't a bigger model. It's an attempt to make the same small local model better at actually DOING things. 🔥 ⚠️ Benchmarks are reported by the NeoHorse team and I want to do these tests myself next.

📰 A 1.56GB local model just took the #1# spot for open models under 4B on Artificial Analysis. Let's run it on ThumbLLM. OpenBMB released MiniCPM5-2B TODAY. Stats 👇 🧠 2.52B dense parameters 💾 Q4_K_M GGUF: just 1.56GB 📚 131K native context ⚖️ Apache 2.0 🦙 llama.cpp 🟢 Ollama 🖥️ LM Studio 🍎 MLX 🤖 Native tool calling + agent training Artificial Analysis Intelligence Index v4.2 ... 🥇 MiniCPM5-2B → 15 Qwen3.5-4B Reasoning → 14* Qwen3.5-9B Reasoning → 15* Granite 4.2 3B → 11 *Qwen scores are estimated by Artificial Analysis. So a 1.56GB Q4 GGUF is landing in the same AA Intelligence Index tier as Qwen3.5-9B. 🤖 GDPval-AA v2 Elo → 831 🏦 τ³-Banking → 21% ⚡ 19K output tokens/task vs 56K for Ling 3.0 Tiny OpenBMB even released the training data and an official DSpark speculative decoding model. No trustworthy local tok/s numbers yet. That is the benchmark I want next. 🔥 Let's run it on ThumbLLM! 🔗 HF /openbmb/MiniCPM5-2B

GPT-5.6 Luna survives the model launch storm — still #1# in the world based on OrcaRouter Composite Index 🐳 The frontier moved fast: Claude Opus 5. Kimi K3. GLM-5.3. Qwen3.8. Yet GPT-5.6 Luna still holds the crown on our Model Leaderboard. Current Top 10: 🥇 GPT-5.6 Luna — 75.0 🥈 Claude Opus 5 — 72.8 🥉 GPT-5.6 Sol — 72.0 #4# GPT-5.4 Pro — 71.7 #5# Kimi K3 — 70.9 #6# GLM-5.3 — 70.0 #7# GLM-5.2 — 69.9 #8# Qwen3.7 Max — 69.8 #9# GLM-5.3 Flash — 69.0 #10# Grok 4.6 — 67.4 This isn't another benchmark beauty contest. Orca Composite Index: 40% Human Preference 30% Independent Benchmarks 20% Production Evidence 10% Ecosystem Adoption Benchmarks measure models in the lab. We measure which models actually win. 🐳

🧭 Can an AI model really teach itself to improve when it's only handed a vague goal, with no task spec and no reward? A team from ByteDance Seed and collaborators built a benchmark to find out. Title: Aspire: Can Models Self-Evolve from Vague Goals? URL: The benchmark hands the model nothing but a vague goal, lets it handle interpretation, training, and verification end-to-end, and then measures real capability gains against a hidden evaluation set. 🎯 Highlight 1: The cost of ambiguity Simply rephrasing a task as vague dropped Claude Opus 4.8's score from 32.90% to 27.07% and GPT-5.6's from 36.23% to 29.58%. Agents burn time figuring out what to optimize for, leaving less time for actual training. 🔁 Highlight 2: Execution isn't the same as improvement In self-evolution runs with Qwen3.5-4B/9B, most final checkpoints underperformed the base model, and 21 of 24 runs were rolled back to their starting state by the safety mechanism. Training against narrow self-evaluations produces local gains that fail to transfer to hidden evaluation. 🛠️ Highlight 3: Harness evolution can't beat the baseline either Even when agents were allowed to rebuild their own execution harness, the best successor harness scored 27.22 versus 28.64 for the reference implementation — falling short due to narrow verification and gaps in output completeness checks. 💡 This research makes clear that deciding what to improve is the next big wall standing between agents and sustained self-improvement. #LLMAgents# #SelfEvolution#

Qwen3.8-2.4T-A95B, compressed two ways at once: 25% of the experts pruned with REAP, and the rest quantized to NVFP4. Even with a quarter of the experts gone and 4-bit weights, GPQA Diamond holds at 91.5 vs 92.6 for the full-precision base. ~99% recovery. Serve on @vllm_project:

Qwen3.8-2.4T-A95B, a 2.4-trillion parameter MoE, quantized with no accuracy loss. MoE layers to NVFP4, attention layers to FP8 block, via LLM Compressor. On GPQA Diamond: 93.1 quantized vs 92.6 for the full-precision base. Full recovery. Serve on vLLM:

Qwen3-8B now has a DFlash speculator! 82.2% first-token acceptance on math reasoning. 3.74 avg tokens accepted per step. Built with the Speculators library. Training compute sponsored by @modal. 🙏

🚀 Our DSpark for Qwen3.8-27B beats native MTP with the same 8 speculative tokens on 4×H200. Up to 52% faster single-stream decoding and 23% higher peak throughput. On 8-needle MRCR, it averages 4.30 accepted tokens beyond 1M context. Model: Demo below, see for yourself!