The full Qwen3.8-Flash-Next on a 24 GB RTX 3090 with 64 GB of DDR4. Strata decodes at 78 tok/s against llama.cpp's 18, and reads prompts about 3x faster.

DDR4-2133 vs 3000 on an RTX 3090 box: no change for models that fit on the GPU, 13 to 29% faster with MoE experts offloaded. Plus why 3200 wouldn't train.

There is no official Qwen 3.8 35B-A3B. I ran empero-ai's distill against Qwen 3.6 35B on an RTX 3090: same score, same speed, a fifth of the reasoning tokens.

Qwen3.8-Flash-Next Coder on a 12 GB RTX 3060 and 32 GB of DDR4. Strata prefills 6-7x faster than stock llama.cpp. One batch setting closes most of that.

A drafter benchmark on this site was Codacus's GTX 1060 figure. Our own 3060 numbers, why it's the CPU and not PCIe, the July audit, and what a 12 GB card should run. Plus 4090 and break-even corrections and seven Quick Hits.

Bonsai 2 27B fits a 12 GB card and decodes 1.8x faster than its Q4, then drops from 19 to 12 of 47 on a routing split, all of it on one field. Plus 23 used-GPU price corrections, the 3060 fair-price ladder, and eight Quick Hits.

Bonsai 2 27B, PrismML's ternary Qwen3.8, on a 3090: 1.8x the decode of Q4 from a 7.2 GB file, then 12 of 47 against 19 on a routing split. One field broke.

Jev picks from options you supply instead of writing, in one pass, with a probability per answer. What it is, what it costs, and the open version for a 3090.

Jev-style parallel decisions in llama.cpp on Qwen3.6-27B, RTX 3090: 1.3x faster at four tokens, 5.5x at fifty, 21 vs 23 of 47. The gain is the tokens you skip.

Codacus's parallel-decision llama.cpp branch answers a schema in one forward pass, and the ten-seeds 47-item split is the ready-made test. Plus ChatGPT-User fetches sliding 5 percent a week while OAI-SearchBot tripled, the v0.4.0 pin on the 3060, and a third bench rig.

Stock llama.cpp put 5.4 GB of a 177B MoE on an RTX 3090 and a 3060 kept pace; -ncmoe 29 buys 34 to 40 percent and a real 32 GB box does 8.7 tok/s. Plus a 4 GB GTX 1650 at 20 tok/s on a 35B MoE, and the MoE primer corrected in public.

A $60-class GTX 1650 4 GB runs Qwen3.6-35B-A3B at 20 tok/s on 32 GB of RAM. The RTX 3060 in the same slot does 28 at the same setting and 39 tuned. Measured.

Qwen3.8-Flash-Next on an RTX 3090 and a 3060. Stock, the 3090 matches the video's 3060. One flag buys 34 to 40 percent. A real 32 GB box: 8.7 tok/s, not 22.

Nvidia PAIR split 20 requests 10/10 across an RTX 3090 and a 3060 for 1.07x, and my own router managed 0.52x. Plus ten LoRA seeds that all landed at or below the base model, and the llama.cpp pin moving to v0.4.0.

Nvidia's new home-network AI router split 20 requests across an RTX 3090 and a 3060 for 1.07x over the 3090, then dropped half a 27B queue. Mine: 0.52x.

Ten LoRA seeds on identical data spread 3.62 points, against 4.65 for prompt compilation. Not one beat the no-adapter base. Measured on an RTX 3090.

A frontier model wrote a skill that made our local 27B 10.6 points better. Nine more compilation runs showed the number was noise. Plus the LoRA substrate nobody has measured, MoE routing traced, and 277 GB in a file with no name.

Compiling a skill into a prompt costs 1,383.9 tokens per call, forever. Putting it in a LoRA costs 7.55 GB and a training run. Only one has been measured.

A frontier model read my logs and wrote a skill that beat my local 27B baseline by 10.6 points. Nine more compilation runs showed the number was fake.

Firsthand A-B-B-A bench of Ornith 1.5-35B-A3B against Qwen 3.6-35B-A3B on one RTX 3090. Generation, prefill, VRAM, and the noise floor under all three.

Two independent labs shipped competitive open MoE models in 2026. I read both technical reports. One of them is built on NVIDIA's architecture, data and tokenizer.

Qwen 3.8-27B spent 14,953 tokens on a line the same file writes in nine. All 164 HumanEval problems measured: 92.8% of output is thinking. Plus the four runs that tie it with 3.6.

I traced every expert routing decision Qwen 3.6-35B-A3B makes across six workloads on an RTX 3060. Routing isn't flat, and 112 slots is the whole answer.

Qwen 3.8 27B generates at full speed on a 3090 and still crawls. All 164 HumanEval problems measured on the chat path: 92.8% of output is thinking.

Firsthand A-B-B-A llama-bench of Qwen 3.8-27B against 3.6-27B on one RTX 3090, same quant publisher. Generation lands within a percent. VRAM costs 254 MiB.

Adding a second RAM stick to a mini PC lifted CPU token generation 52-58% across four models. Prompt processing moved under 2%. Measured before and after.

The registry was patched and rebuilt by 6 July. On 8 July the agents had a new board, encoded in directory names. Plus the ranking null behind it.

Every offloaded layer on a 3090 costs about half a millisecond, with no knee anywhere in the curve. Firsthand sweep, the two rules it broke, and a new 26B MoE.

Eight engines that stream MoE experts from disk appeared in four months. Three have no license file at all. Here's the verified state of each.

One model, three places its experts can live: VRAM, RAM, SSD. We measured the first two on Gemma 4. TurboFieldfare just added the third.

1,238 AI staff asked Washington to slow the frontier. Every mechanism proposed works at release — the only moment an open release can be touched.

The whole 35B sits on a 24GB card with 2.4 GiB spare, no expert offload. Firsthand numbers, the harness caveat, and why max offload loses to a 3060.

GPU prices are spiking due to GDDR7 shortages and AI datacenter demand. Here's what's happening, which cards are affected, and strategies for local AI builders.

I measured Qwen3.6-35B-A3B on a 12GB RTX 3060: 38 tok/s, flat through 8K, via llama.cpp --n-cpu-moe. Why expert offload flies where dense models die.

Two trillion-parameter 'open' models dropped in ten days — Qwen 3.8 and Kimi K3 — and you can't run either. The same week, Hugging Face got breached and its own responders, blocked by commercial-model guardrails, ran the forensics on GLM 5.2, an open-weight model on their own hardware. Same lesson f

Hugging Face says an autonomous AI agent breached its internal infra on July 16. The models you download are safe — here's what was and wasn't hit.

Qwen 3.8 (2.4T) and Kimi K3 (2.8T) both went 'open' in ten days. Neither fits your GPU. Here's Qwen's real open-weight cadence and what to run today.

DeepSeek closed a ~$7.4B round with China's state AI fund holding the only voting rights, funding open weights like national infrastructure. The same week, Demis Hassabis called for a FINRA-style body to screen frontier models before release, open or closed. Plus: the twist where the country bankrol

We rented an A100 and an H100 back to back: the pricier card cost less per training run because it finished in half the time. Plus honest notes from the AI Engineer World's Fair floor, and a Qwen 3.7 open-weights status check.

Qwen's own release cadence says the 3.7 open weights should already be out, and they're not. Plus GLM-5.2 running locally: a frontier open model that takes serious hardware.

GLM 5.2 is 753B params and 1.51TB at full precision. Run it locally: the live Unsloth quant ladder, every GPU and RAM path, and the quant to actually target.

Qwen split into a closed frontier (Max, Plus, VLA) and an open mid-tier (3.6-27B and 35B-A3B). The 3.7 open weights aren't here yet. The honest read.

DeepSeek V4 is going from 'just dropped' to 'actually deployable' as the tooling catches up — plus a July 24 deprecation that breaks your code if you're not watching, and a 4x price cut on V4 Pro.

Ollama 0.30 quietly changed how Apple Silicon runs models, auto-routing safetensors to MLX and GGUF to llama.cpp Metal. The open workhorses are now Qwen 3.5 9B and Qwen 3.6 27B. Plus: Qwen's last two flagship models shipped closed.

Ollama 0.30.0: llama.cpp integration, flash-attention default for Qwen/Gemma, broader model support. Firsthand upgrade notes, known issues to watch.

MiniMax M3 ships with frontier benchmarks but no downloadable weights yet. The Windows unified-memory hardware shift is coming for Apple Silicon's lead. And a personal note about who I'd like to see at AI Engineer World's Fair.