AI Coding Live News — Real-Time 𝕏/Twitter Updates & Breakthroughs

👀123

🔥96

NVIDIA open-sources Nemotron 3.5 Super VL (120B-A12B multimodal agent model, 1M context) and the 24GB 2-bit local Bolt 74B-A9B

On 2026-10-15 NVIDIA released Nemotron 3.5 Super VL, an open multimodal agent model with 120B total / 12B active parameters: a Mamba-2–Transformer hybrid MoE with a C-RADIO vision encoder, text/image/document/video input, up to 1M context, and built-in MTP plus LiLiCorr speculative decoding. Weights, datasets and recipes ship under OpenMDW-1.1 on Hugging Face, OpenRouter, ModelScope and NIM. Alongside it, NVIDIA published Bolt 74B-A9B 24GB, a natively 2-bit (QAT) local agent model for RTX/DGX Spark; the 4GB Bolt E4B follows on Oct 23.

Key Takeaways

Scale: Super VL is 120B total / 12B active with up to 1M context; BF16 reference serving needs 4×H100 or 2×GB200, with an NVFP4 build for smaller footprints (model card)
Coding (vendor-reported): SWE-bench Verified 71.10, Terminal-Bench 2.1 66.70, LiveCodeBench v6 88.22; in the same table Qwen3.5-122B-A10B scores 71.9 / 48.70 / 76.7 and GLM-5.3-Flash 79.10 / 82.70 / 80.6
Agent evals: Hermes Bench 63.44% (Nemotron 3 Ultra 56.33%), LangChain Evals 87.05% (Ultra 84.3%); NVIDIA claims up to 5x single-GPU throughput vs open peers and 600+ TPS on Baseten (tech blog)
Bolt 74B-A9B 24GB: native 2-bit (IQ2 + QAT), 256K context; Terminal-Bench 2.1 50.98 and BrowseComp 66.75 vs 44.38 / 48.74 for Qwen3.6-35B-A3B; PinchBench 87.82 is slightly below Qwen3.6's 88.07 (Bolt card)
Serving: LiLiCorr speculative decoding gives up to 2.3x faster generation on DGX Spark; the Dynamo recipe serves 8x more users with 85% lower TTFT on 16×B200 at 50 tok/s/user

Read full story

Linked Models & Agents:

⚡ nemotron-3.5-super-vl⚡ nemotron-3.5-bolt-74b-a9b⚡ nemotron-3.5-bolt-e4b⚡ Nemotron 3 Ultra🤖 hermes-agent 🤖 opencode 🤖 OpenHands (All-Hands AI)🤖 Cline (VS Code Agent)🌐 Nvidia (NIM)

#Nemotron 3.5#NVIDIA#MoE#Mamba#Agents#Hugging Face#OpenRouter

Fish Audio@FishAudio·Collected 22d ago

👀206

🔥93

View on 𝕏 ↗

Fish Audio ships Drama 3: open-weight 35B MoE TTS with word-level emotion direction, #1 on Artificial Analysis Pronunciation Robustness at 93.2%

On 2026-10-13 Fish Audio launched Drama 3, a TTS model directed with plain-language bracket tags, open weights on Hugging Face, and API header model: drama-3. The backbone is a ~35B MoE (~3B active per token, initialized from Qwen3.5-35B-A3B's language tower) with a codebook-MTP audio decoder and preference-model RL. Independent Artificial Analysis Pronunciation Robustness: 93.2% vs 88.1% for S2.1 Pro; 83 languages and up to 16 speakers; Community License free for commercial use under US$500k annual revenue. Launch promo through Nov 12: about $10 per million characters on the API.

Key Takeaways

Open weights on Hugging Face fishaudio/drama-3 under Community License v1.0 (free commercial use under US$500k ARR; no non-consensual voice cloning)
Architecture: 35B MoE / ~3B active; hybrid attention (1 full-attn layer per 4); codebook-MTP decoder over 10 codebooks; initialized from Qwen3.5-35B-A3B LM
Benchmarks: Artificial Analysis Pronunciation Robustness 93.2% (#1) vs S2.1 Pro 88.1%; brand/place/tech names 95.9%, context-dependent readings 95.3%
Vendor H200 streaming: 271 frames/s single stream; concurrency-1 TTFA p50 88.9 ms; concurrency 16 ~927 frames/s, RTF 0.373
Ship: API POST /v1/tts with model: drama-3; self-host via SGLang-Omni (sgl-omni serve --model-path fishaudio/drama-3, typically ≥48GB NVIDIA VRAM)

Read full story

Linked Models & Agents:

⚡ Drama 3⚡ Fish Audio S2.1 Pro⚡ Qwen3.5-35B-A3B🌐 Fish Audio 🌐 Artificial Analysis 🌐 SGLang

#Drama 3#TTS#open weights#MoE#Fish Audio#SGLang-Omni#voice cloning#Pronunciation

JetBrains@JetBrains·Collected 28d ago

👀172

🔥94

JetBrains ships Mellum2.1: 12B MoE (2.5B active) open coding-agent model, SWE-bench Verified jumps 2.0→47.0

On Oct 8 JetBrains released Mellum2.1 Thinking: same 12B MoE / 2.5B active / 131K / Apache 2.0 architecture, with gains almost entirely from large-scale RL in real sandboxes. Self-reported same-pipeline scores: SWE-bench Verified 47.0 (vs Mellum2 2.0), LiveCodeBench v6 82.0. On Hugging Face; vLLM serve now; GGUF/MTP coming soon.

Key Takeaways

Spec: 12B MoE / 2.5B active, 131K context, Apache 2.0; architecture unchanged vs Mellum2 (HF card)
Agentic jump (same pipeline): SWE-bench Verified 2.0→47.0, SWE-bench Pro 0.0→28.0, Terminal-Bench 2.1 0.6→17.4 (Pi v0.73.1)
Coding: LiveCodeBench v6 82.0 (leads Qwen3.5-9B 75.4), HumanEval+ 91.5, BFCL v4 62.3
Speed: fastest under load in JetBrains’ group; MTP ~1.6× single-request; still trails Qwen3.5-9B on SWE Verified 50.0 / Pro 38.0
Ship: vllm serve JetBrains/Mellum2.1-12B-A2.5B-Thinking --reasoning-parser qwen3; GGUF/Ollama/LM Studio/MTP head coming (blog)

Read full story

Linked Models & Agents:

⚡ Mellum2.1⚡ Mellum2⚡ Qwen3.5-9B⚡ Gemma 4 E4B🤖 Pi 🌐 JetBrains

#Mellum2.1#JetBrains#Open weights#Coding agent#MoE#SWE-bench

ggml-org@AICoder·Collected 31d ago

👀196

🔥84

llama.cpp v0.6.0 ships GLM-5.3-Flash (320B) and Qwen4Exp MTP speculative decoding, a new llama_batch_ext API, and up to ~3x faster few-row mat-mul on Apple GPUs

ggml-org released llama.cpp v0.6.0 on Oct 5 (v0.5.0 was Sep 23). It brings GLM-5.3-Flash (GLM5-Next, a 320B text+vision hybrid MoE) into a stable tag, adds MTP speculative decoding for Qwen4Exp (~1.5x decode on DGX Spark), introduces the llama_batch_ext / llama_process() API for mixed token and embedding batches, adds few-row MMA mat-mul kernels on Metal (up to ~3x for speculative and batched decoding), and bumps ggml to v0.26.0. Session file formats are bumped too.

Key Takeaways

New models: GLM-5.3-Flash (GLM5-Next, 320B KDA/DSA hybrid text+vision MoE), Clef decision model, Ling 3.0 VL, LFM2.5-Encoder
Speed: Qwen4Exp MTP speculative decoding ~1.5x decode on DGX Spark; Metal few-row MMA mat-mul up to ~3x
API: llama_batch_ext + llama_process() for mixed token/embedding batches; server and examples migrated
Compat: LLAMA_SESSION_VERSION 11 and LLAMA_STATE_SEQ_VERSION 4, so old session/state files must be regenerated
Server: /v1/models reports modalities, /v1/embeddings takes typed vision/audio/video input, new /v1/systemone endpoint

Read full story

Linked Models & Agents:

⚡ GLM-5.3-Flash⚡ qwen4exp⚡ clef🌐 ggml-org 🌐 zhipu

#llama.cpp#Ggml#GLM-5.3-Flash#Qwen4Exp#MTP#speculative decoding#Metal#local inference#Release

Brian@Brjen·Collected 47d ago

👀84

🔥87

View on 𝕏 ↗

RTX 3090 local coding-agent bake-off: Qwen3.8-27B Q4 + llama.cpp ~59 tok/s with reliable tools

@Brjen’s team spent a week comparing local coding-agent stacks on one RTX 3090 (24GB). Winner: Qwen3.8-27B Unsloth UD-Q4_K_XL with llama.cpp (65k context, q8 KV, native MTP depth 2, fully on GPU). Real multi-step coding runs hit ~59 tok/s (~46 tok/s at 45k context) and passed nested tool JSON, fixing a failing test, and a clean unified diff. They rejected Q3 for bad tools, heavier quants for context/MTP tradeoffs, and Flash-Next setups that needed CPU offload.

Key Takeaways

Winner: Qwen3.8-27B UD-Q4_K_XL + llama.cpp, 65k context and MTP, fully on a 3090.
~59 tok/s on multi-step coding; passed tool JSON, test-fix, and correct diff gates.
Q3, heavier quants, and offloaded Flash-Next were less reliable on 24GB.

Read full story

Linked Models & Agents:

⚡ Qwen 2.5 Coder 32B

#Qwen3.8#llama.cpp#Local#coding agent#RTX 3090#Unsloth#MTP#Benchmark

Unsloth AI@UnslothAI·Collected 62d ago

👀76

🔥91

View on 𝕏 ↗

Unsloth: GLM-5.3-Flash local GGUF inference up to 3.3× faster with MTP

Unsloth says optimized local GGUF decoding makes GLM-5.3-Flash up to ~3.3× faster (typically ~1.6–3.4×), with multi-token prediction (MTP) and faster long-context decoding. 3-bit quants still target ~128GB setups via Unsloth Desktop or llama.cpp; docs and Hugging Face GGUF weights are updated.

Key Takeaways

Speed: local GGUF inference up to ~3.3× (typical ~1.6–3.4×).
Features: MTP enabled plus faster long-context decoding.
Run: 3-bit on ~128GB via Unsloth Desktop / llama.cpp; docs and HF GGUFs live.

Read full story

Linked Models & Agents:

⚡ GLM-5.3-Flash 🌐 Zhipu AI (智谱 AI)

#Unsloth#GLM-5.3-Flash#GGUF#llama.cpp#MTP#local inference#Z.ai

vLLM@vllm_project·Collected 67d ago

👀56

🔥95

View on 𝕏 ↗

vLLM 0.28.0 Released: Full-Stack Kimi-K3, DeepSeek-V4 Sparse MLA, and Tiered Disk KV Cache

vLLM 0.28.0 landed with 584 commits, bringing stack-wide Kimi-K3 support, end-to-end DeepSeek-V4 sparse MLA with DFlash2/DSpark speculative decoding, Model Runner V2 disaggregation, and tiered disk KV cache.

Key Takeaways

Adaptive speculative token budgets slash end-to-end TTFT by 55–65% with 1.5–3x faster sequence-parallel kernels;
End-to-end DeepSeek-V4 sparse MLA and Kimi-K3 shared-expert sharding save ~17 GiB VRAM per GPU;
Tiered KV cache introduces a disk tier, doubles max batched tokens to 16,384, and supports diverse hardware backends.

Read full story

Linked Models & Agents:

⚡ DeepSeek-V4-Flash ⚡ DeepSeek-V3

DeepSeek@DeepSeek_AI·Collected 67d ago

👀47

🔥94

View on 𝕏 ↗

DeepSeek Open-Sources Multi-Token Prediction (MTP) Agent Harness Framework: 300% Multi-File Generation Speedup

DeepSeek open-sources its Multi-Token Prediction (MTP) agent acceleration framework. By speculatively predicting syntax blocks in parallel, multi-file code diff and test generation in Cline and Aider achieves over 3x speedup with zero degradation in AST validity.

Key Takeaways

DeepSeek releases Multi-Token Prediction (MTP) framework tailored for coding agents;
Delivers 300% throughput acceleration during multi-file refactor and test suites;
Directly compatible with local vLLM and Ollama deployments.

Read full story

Linked Models & Agents:

⚡ DeepSeek-V4-Pro ⚡ DeepSeek-V4-Reasoner 🤖 Cline (VS Code Agent)🤖 Aider 🌐 DeepSeek

#DeepSeek#MTP#Open Source#Agent#Inference