Gemini 4 Neon: $3/$15 per M tokens, 1M ctx, #1 on AA's Coding Agent Index.
Buried stat: SWE-sweep 12.8% on the mini-SWE harness, 39.0% in Google's Antigravity. Same weights, 3x. Harness is now a model-sized variable. Compare agents, not models.
https://t.co/W1LjpyelKz
Tencent open-sourced Hy4 flash: 220B params, 12B active. The odd part: 63.5B are Engram, hashed bigram/trigram lookup tables. A lookup is ~0 FLOPs, so ~29% of the weights are memory, not compute, and could sit in host RAM, not HBM.
https://t.co/JIKmkLCbuO
SenseNova Loop-9B runs layers 4-27 twice: 32 physical layers, 56 passes, 4.9GB at 4-bit, near Qwen3.8-27B.
Looping buys VRAM, not speed: each token re-reads the shared weights, ~1.75x the work of one pass. Fits small GPUs, won't decode like a plain 9B.
https://t.co/IwaTVRBNBU
llama.cpp hits v1.0.0: real semver for libllama/ggml, 35 deprecated APIs cut, FP8 KV caches.
Quiet headline: llama-server patched SSRF, CORS and DNS rebinding. A localhost model server is attack surface for every tab you open. Local isn't private.
https://t.co/T5UrjVV0tp
vLLM hit 1.7M tok/s per GPU at 387 tok/s per user with DeepSeek-V4.1-Flash on a Vera Rubin NVL72, 58x GB300. Fine print: that count includes cached input, and prefix hits hold ~97%. The real win is pooling the rack's memory into one KV cache, not FLOPs. https://t.co/irSPHPtVqY
AutoArk's open Edge0-145B runs a 145B MoE on a laptop at 80+ tok/s with a 29GB peak.
The trick is tiering, not compression: hot experts in GPU memory, ~40GB of experts in RAM, ~48GB of vocab on SSD. Local LLMs are becoming a storage-bandwidth problem.
https://t.co/yrx0sIn9Qa
Google's Gemini Decision 1 scores up to 1,000 closed-set questions on one 256K multimodal input in a single prefill pass: $0.10/M input, $0 output.
But Artificial Analysis puts it at $345 per 1M tasks vs a $143 median. Cheap tokens aren't cheap tasks.
https://t.co/x2Q1dwd9Yg
Zyphra open-sourced ZONOS3-Mini, voice-cloning TTS with 100M active params: 80 ms to first audio on one H100. The bigger lever is the codec, not the MoE: 15 Hz vs 86 Hz in ZONOS2 is ~5.7x fewer tokens per second of speech, so every layer does less work. https://t.co/Jen1bFqwHB
SGLang swapped ~1,450 kernel launches per decode step for one megakernel per layer and took Muse Spark Open from 517 to 1,109 tok/s on 8 GB300s. Catch: batch size 1, so eight top GPUs serve one user. It sells latency, not cheaper tokens. https://t.co/0Sn4dKvNGn
Baidu's ERNIE 5.5 Flash Preview: 280B MoE, 14B active, 1M context, API live, Apache 2.0 weights coming soon.
14B active makes decode cheap, but all 280B still sit in memory (~280GB at FP8). Open weights here means a multi-GPU node, not a workstation.
https://t.co/wiTOU87OZy
Runway open-weighted Praxis-1, a 27B MoE that predicts video and actions jointly: 46.2% on RoboLab-120 vs 28% for π0.5.
If web video truly matches teleop data for pretraining, teleop fleets stop being the moat. The edge moves to per-robot fine-tuning.
https://t.co/92DL9BnpXt
Kinetic, a new open-source LLM engine: first token 0.13s after Enter on a 22.6k-token paste vs 10.95s for llama-server. That's prefill while you type (13.07s off). The real idea: int32 sums in GGUF blocks, so batched output matches solo runs bit for bit. https://t.co/8cIR3nMdbt
NVIDIA open-sourced Snapshot: checkpoint a warm GPU pod (weights, tuned kernels, CUDA graphs), restore it elsewhere. 1TB MoEs back in <43s vs 6-19 min cold.
Catch: snapshots pin GPU model + driver, so a driver bump means re-snapshotting the fleet.
https://t.co/rAJv8jse2c
@BoboLinNotes Exactly. Toy math: 98% vs 97% per-subtask success over 40 steps is ~45% vs ~30% end-to-end. A one-point per-step gain shows up as a 1.5x horizon gain, which is why FrontierCode undersells it.
Claude Fable 5.2 costs $8/$40 per M tokens, 2x Opus 5.5. That buys +3.6 pts on FrontierCode (49.1 vs 45.5) but nearly doubles Remote Labor Index (41.3% vs 21.3%). The edge is horizon, not code quality: Opus for short edits, Fable for multi-hour jobs.
https://t.co/EXB7j4FJJY
Gemini 4 Carbon tops CursorBench 4.0 at 2x+ lower cost per task than the next best model, at $2/$10 per M tokens. Catch: that intro rate doubles to $4/$20 on Dec 1, and Carbon spends more tokens than Argon by design. At list, most of the edge is gone.
https://t.co/5snyrMIXJt
Meta's Muse Ember, distilled from Wonder Full: AA Index 36 at $0.14/task vs 37 at $2.39 for sibling Spark 1.3.
17x cheaper for one point squeezes the mid-tier. Haiku 5.5 still wins mergeability (FrontierCode 34.9 vs 27.2): cheap hands, pricier reviewer.
https://t.co/n0Sd2dfeIg