Got continuous batching working with SSMs in mlx-lm. Here's... four OpenCode agents simultaneously running Nvidia's Nemotron Nano on 64GB M4 Max. This is a nice model for smaller machines since it's MoE + hybrid attention (small cache).show more

I tested MTPLX v2 with QWEN 3.6 27B and... compared it with oMLX without cache on M5 Max and DGX Spark on vllm using nvfp4 model version. More details in 🧵 I've reached 82.8 tps of max decoding speed! 🔥 Custom Metal Kernel design specifically for this model and for Apple Silicon is just perfect! This is the way forward! Great job Youssof Al Toukhi Look at the website here! 👇 Here a website with recap, built with GLM 5.2 running locally 💪 First chart and preview from the website.show more

Got DeepSeek V4 Flash running on 2x H200s at... 160–200 tok/s on JarvisLabsAI this speed is perfect for in the loop things I do with these agents. It feels like a GLM-5.2-class model with much lower hardware requirements. I’m going to daily-drive it for a bit and see how it performs, but first impressions are really good.show more

Day 11/90 of Inference Engineering How does vLLM work... and how is it used in production? Before we discuss how vLLM works internally, it helps to understand what vLLM is. At a high level, vLLM is an inference engine that is designed to serve LLMs to thousands of concurrent users efficiently while managing scarce compute and memory. The goal for vLLM is to maximize throughput and minimize latency; optimizing for the best inference economics and experience for end users. With every request from the end user, it eventually ends up in the engine core, gets scheduled alongside other requests from other concurrent users, executes on the GPU, and updates the KV cache with the new key and value vectors, and streams the tokens back to the user. The Scheduler decides what requests should execute next while continuously batching requests together to maximize GPU utilization. Continuous batching is an inference optimization that allows new requests to join a running batch as other requests finish generating tokens. This helps with keeping the GPU utilization high instead of letting it sit idle waiting for an entire batch to complete generating. After the scheduler dispatches the selected batch to the Model Executor, the Model Executor prepares the tensors and metadata required for inference, retrieves each request’s block table from KV Cache Manager, launches the optimized transformer forward pass on the GPU, computes the logits, updates the KV cache with the new key and value vectors, and finally returns the results for sampling and streaming. The KV Cache Manager uses the PagedAttention memory layout to allocate fixed-size cache blocks on demand and maintains a Free Block Queue on the CPU that tracks which blocks in the GPU’s Paged KV Cache are currently free. When a request needs additional KV cache space, the KV Cache manager takes a free block from the queue and assigns it to that request, thus avoiding an expensive search through GPU memory for available cache blocks. All of these components form the core of vLLM’s inference engine. The Scheduler determines what requests are executed, the Model Executor determines how those requests are executed, the KV Cache Manager determines where each request’s KV cache lives using the PagedAttention Memory Layout. This architecture enables vLLM to serve thousands of concurrent requests with high throughput, low latency, and efficient GPU memory utilization. Heres a little animation that visualizes everything! - I've also completed the forward pass for my mnist.c project. I had a nice chat with shrey birmiwal, such a knowledgeable guy. Excited to learn more about vLLM and implement a tiny-vLLM one day.show more

Before the week ends, let's acknowledge one of the... most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: 🧠 LLMs → NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. → Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. → StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. → Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. → JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. 🎨 Image gen (the surprise of the week) → Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. 🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped) → Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. 👁️ Vision & VLMs → PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. 🎬 Video, 3D & World Models → NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).show more

Max B Released From Prison After 16 Years After... being locked up for over 15 years, Max B is finally free. He's hitting the ground running. He already linked up with French Montana, popped up on the field at the Jets game, and he's got a welcome home party in NYC tonight, followed by a string of other events on the East Coast. "I got one mission: to restore musical excellence," he told Okayplayer this week. "It's brilliant. That's all I'm going to tell you."show more

Gemma 4 12B QAT (dense) achieves 1000+ tokens/sec prefill... on 8GB VRAM with 120k context Gemma 4 12B QAT (dense), TurboQuant (Without MTP), RTX 4060 8GB VRAM: Prefill: 1000+ tok/s (42% increase) Decode: 25+ tok/s (25% increase) Context: 120k (150% increase) prefill was 700 tok/sec and decode 20 tok/sec with only 48k context without turbo quant (older test with mtp link in the comments) llama.cpp TurboQuant flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -c 120000 --cache-type-k q8_0 --cache-type-v turbo3 -ngl 99 --port 8080 tested with a 27k prompt, 120k context loaded. -ngl 99 here isn't a typo, full 12B dense, every layer on GPU, on an 8GB card. that's the part worth sitting with. The model has vision, audio input, thinking/reasoning and fits your 8GB card. TurboQuant's KV cache savings are what free up the room to do that at 120k context. side by side with yesterday: 26B A4B MoE got 320+ tok/s prefill. this dense 12B is clearing 1000+ rig: RTX 4060 8GB · i7H · 16GB RAM same two flags as yesterday, different model size: --cache-type-k q8_0 --cache-type-v turbo3 thanks to TheTom/llama-cpp-turboquant, TurboQuant fork of llama.cpp by Tom Turney (Tom Turney) to make this work. unsloth's model quant huggingface and the llama.cpp fork github link in the comments Do you prefer a dense or a MoE for your 8GB card?show more

Been using Qwen 3.8 27B (Q4) locally on 64GB... of VRAM. Here is the verdict: SLOW 18 tps with ZERO system prompt to process and that degrades significantly with a harness system prompt and as the context window grows. RIP if you have to compact. I had it implement this PRD and it's been running for 6 hours. By comparison Grok 4.6 and Kimi K3 hosted finished in about ~30 minutes. High hopes, but these 27B variants are too dense. This is not a consumer grade local model - and I consider consumer grade to be anything up to $5000.show more

Run Gemma 4 26B MoE on 8GB VRAM with... 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the repliesshow more

AI in robotics gets all the attention right now,... but sometimes the most interesting work is very practical. Viet built a small vision system that counts potatoes on a conveyor belt. No giant dataset. No huge model. Just a clear problem and a smart setup. He used Ultralytics’ ObjectCounter, trained a tiny YOLO11 nano model, and because there was no potato dataset, he annotated a single frame with SAM 2 and trained from that. One frame. Still works across the whole video. It is a good reminder that useful AI in industry often looks like this. Focused. Lightweight. Solves a real task. If you work in manufacturing or robotics, these small systems are usually the fastest wins. They save time, reduce errors, and do not need massive infrastructure. Nice work, Viet. His projects: —- Weekly robotics and AI insights. Subscribe free:show more

IT'S MONDAY MORNING. YOU'RE ASLEEP Your trading agent has... already opened three positions, adjusted risk exposure twice, and closed a profitable trade. Before you reached for your phone. $100,000 simulated capital. 72 hours of full autonomy. $6,000 prize pool. Zero coding required > no-code agent builder, if you can describe how you think about markets, Bloome builds the agent > four specialized agents running simultaneously: research, technical analysis, risk management, execution > scoring rewards discipline, a 7% return with clean risk management beats a 14% return with chaos > 1,000 invite slots, competition activates at 50 participants, the field is still small right now > first place: $3,000 · second: $2,000 · third: $1,000 · payouts go out June 12 this isn't a trading bot competition it's a test of whether you can think systematically about markets and encode that thinking into an autonomous system the old model: sit at a screen, make emotional decisions, compete against algorithms that never sleep the new model: build the algorithm → let it run → go to sleepshow more

We are in an insane run of open-weight drops.... Every modality, open source is winning. This is what an open source AI summer ☀️ looks like: 🧠 LLMs & Reasoning → DeepSeek-V4-Flash-0731 (my king 👑): 304B MoE refresh, Terminal-Bench 2.1 jumps 61.8→82.7 over the preview, DeepSWE 7.3→54.4. Closes in on Opus-4.8 on Agents' Last Exam (25.2 vs 25.7). MIT. → Muse-Glimmer-30B, from Meta (they are back!!): their first open agentic model. ~29.6B dense + perception encoder, 131k+ context, built to run fully local, no cloud. Apache 2.0. → Liquid AI LFM2.5-2.6B: 2.69B params, 131k context, 220 tok/s on an M5 Max in under 2.5GB RAM. Competitive with models 4x larger on agentic tasks. → inclusionAI Ling-3.0-flash: 124B total, only 5.1B active, ~12% the size of their old 1T flagship Ring-2.6, matches it on key benchmarks. MIT. → inclusionAI Ling-3.0-tiny: 7.9B total, 1.3B active, 86-90 tok/s on an M4 Pro MacBook at ~8GB peak memory. MIT. → NVIDIA Nemotron-3.5-Lightning-30B-A3B: hybrid Mamba-2+MoE+Attention, up to 1M context, runs on a single H100 or DGX Spark, SWE-bench Verified 52.8. → deepgrove maple-preview: 20B-A1B ternary-weight reasoner, 218 tok/s on a Mac mini M4, 5.3GB checkpoint. MIT. → BigBang-v1 (endless-frontier): fine-tuned from Qwen3.6-35B-A3B via a self-evolving generator/critic synthetic-data loop. Lands aggregate performance between DeepSeek V4 Flash (284B) and V4 Pro (1.6T), at 35B. Apache 2.0. 🎬 Video → MiniMax-H3: 33B dense omni model, native stereo audio, up to 2K/15s. 3.6k+ likes already. → Minimax-H3-Turbo (lightx2v): Apache-2.0 turbo distillation of H3 for fast inference. → Lightricks LTX-2.5: image-to-video update, custom Gemma-4-12B text encoder, a markedly stronger distilled model. 🔊 Voice → NVIDIA NemotronLabs VoiceChat-11B: full-duplex speech-to-speech, ~450ms turn-taking, #2 on open VoiceBench, and the first open full-duplex model with live tool-calling mid-conversation. 🛡️ Safety → Mistral Shieldstral-1.0-3B: 3B multimodal guardrail that takes your safety policy as plain text instead of fixed categories. Beats LlamaGuard-4-12B and ShieldGemma-9B on HarmBench (99.4) and ToxicChat (84.1) at a fraction of the size. Apache 2.0.show more

A CHINESE GUY PUT 4 MINISFORUM MS-S1 MAX MINI... PCs IN HIS BEDROOM AND TURNED THEM INTO A 24/7 AI AGENT CLUSTER. TOTAL POWER BILL: ABOUT $44/MO. each box is a tiny local AI workstation built around the Ryzen AI Max+ 395. around $3,000 per unit gets him 128GB of unified memory, 2TB storage, dual 10GbE, and up to roughly 96GB usable as VRAM on Linux. one MS-S1 Max can already run serious open models without touching the cloud. Qwen3-Coder 30B for fast coding, Llama 3.3 70B for heavier reasoning, and larger research models overnight when speed matters less than free inference. four boxes in one room changes the whole game. he is not opening a chatbot, paying for every loop, or shutting agents down before sleep. this is private infrastructure that keeps working even when he is offline. the agents can sort inboxes, review code, summarize documents, monitor feeds, prep meetings, and read papers overnight. on cloud APIs, that kind of always-on stack can easily burn $800 to $1,200 a month if used aggressively. his setup is roughly a $12,000 hardware spend, but the monthly cost is basically electricity. a rack, a switch, a NAS, a small monitor, and four tiny MS-S1 Max boxes turning a bedroom corner into a private inference factory. this is what AI looks like when it stops being rented and starts becoming something you own.show more

you're paying $20/mo for something your $500 GPU can... already do. Gemma 4 26B A4B QAT MoE + Hermes Agent running on a single RTX 4060 (8GB VRAM). Built a vision capable, 100% free, 100% local, private AI assistant that lives in my Chrome browser. No API keys. No cloud. No subscriptions. 100% vibe coded. 0% handholding. It has full context of whatever's on my screen can answer questions, summarize pages, extract data, and see images. Same local model handles everything, no external calls, ever. keep reading for the model and hermes agent tips i learnt while building this locally. Here's the exact setup for anyone running local LLMs on 6-8 GB VRAM: llama.cpp server flags (on my NVIDIA RTX 4060 8gb VRAM): -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --cache-type-k q8_0 --cache-type-v q8_0 -c 150000 --port 8080 Throughput with quantization: Prefill: 200-250 tokens/sec Decode: 20-25 tokens/sec reduce context if oom on 6 gb vram card. Key learnings: - Quantize KV cache to q8 for faster prefill/decode. Prefill goes from 100-150 (unquantized) to 200-250 tok/s (q8). - But watch out, once actual context grows past ~50k tokens on high entropy workloads, q8 KV quantization can cause hallucinations. Low entropy workloads are mostly unaffected. If you see it happening, drop the quantization. This is common across all local models. - In Hermes Agent settings -> Memory & Context, bump compression threshold from default 0.5 to 0.7. Default triggers way too frequent context compression and eats time. Up next: add persistent memory, web search, tool calling, streaming output and whatever you suggest. Running a 26B MoE with vision + 150k context window on 8GB VRAM would've sounded impossible 6 months ago. Works the same on the NVIDIA RTX 3060 Ti, 3070, 4060 Ti, 5060, 2080, or any 8GB card. VRAM is the only requirement. Local AI agents are closer than people think. You just need to know where the knobs are. Model's Unsloth quant hugging face link in the comments. Have you tried Hermes agent by Nous Research yet? What are you building with local LLMs? Drop it below, let's see what this community is shipping.show more

“Fire to the oligarchy!” implying the entire network of... the sole oligarch in Georgia, the dictator himself. Day 70 of large-scale, continuous protests in Georgia. Day 100 overall since the rigged elections. Day 337 of various forms and stages of constant resistance against the regime, since their second tabling of the Russian law on April 3, 2024. The new elections demand is rapidly transforming into “finish with the regime” and most of our partners still have not voiced formal support for new elections that stems from four factors: fraudulent elections; U-turn in foreign policy, in a plot twist to the regime’s own supporters; systemic torture of citizens; an ongoing crisis that will not end. Democracies and even hybrid systems schedule elections for a fraction of this, don’t they? 📷 Nestan Nene Kvinikadzeshow more

3 weeks since ml-intern launched and we just hit... 1M messages exchanged. that's 3.3 agent-years of ML research in 21 days. 2 months worth of research every day. 17,383 training jobs total. talk about AI acceleration. here's some of what people built: Carlos Miguel Patiño replicated the full DeepSeek v4 architecture and pre+post trained a 100M MoE from scratch. → it landed a third place submission on Keller Jordan optimizer competition. autoresearch on SOTA territory. Lewis Tunstall Got the intern to convert Alec Radford's cool new talkie-lm 1930 model to work with transformers. tokenizer, chat template, model conversion etc all one-shotted by ml-intern. someone created entire PhD dissertation chapter on context-aware agentic cyber defense drafted with 16 research subagents. and someone used it to crack an Paul Jankura kernel optimization take-home. (we don't know how to feel about this one 👀 ) just getting started →show more

New open-source agent harness just landed! I got early... access to TrueForge by TrueFoundry and have been running it locally for the past few days. The harness layer deserves as much attention as the model, and open source matters here because you can inspect the loop, run it on your own infrastructure, and swap to the latest or cheaper models. TrueForge handles the runtime work that makes an agent reliable. It drives the tool-calling loop, manages context, coordinates subagents, and executes code in a sandbox, with any model you choose. Every tool call re-sends the growing context to the model, so in practice the harness controls most of what an agent costs to run. A few things stood out from my testing and their published benchmarks. Vendor-Neutral by design. It runs OpenAI, Anthropic, and Google models alongside open-weight models like Kimi, GLM, and DeepSeek. Model routing is a setting, and you can send each task to the model that fits it. On a 14-task enterprise agent benchmark, it matched the accuracy of Claude Managed Agents running the same Opus 4.8 model at roughly 30% lower cost per run (3.8M tokens vs 10M for the same answers). Routing the same tasks to GLM-5.2 held accuracy and brought cost down by about 75%, around $3 per run instead of $12. Fully self-hosted and Open Source (MIT License). I had it running locally with one command, with sandboxed code execution working out of the box. It's time to own your agent harness. Thanks to TrueFoundry for partnering on this post.show more

🚨 Anthropic committed up to 1M TPU chips for... Claude. Openai is leasing TPUs for chatgpt inference. Here's How kernels work on TPUs (deep dive 2/6 by emi) pallas is Google's answer to kernel writing. a python kernel SDK built on JAX. still very experimental (jax.experimental.pallas). on TPU it compiles through mosaic; on GPU it lowers to triton. if you know CUDA, the syntax will feel familiar but the execution model is completely different. in CUDA, grid=(4,4) launches 16 blocks running simultaneously across SMs. in pallas, those 16 iterations run one after another in lexicographic order. no threads. no warps. no blocks. no occupancy tuning. a TPU is a sequential machine with a very wide vector register — more like a CPU than a GPU. performance comes from width: a 128x128 systolic array doing matmul and an 8x128 SIMD vector unit doing everything else. maximum parallelism on chip: 2, one per TensorCore in megacore mode. three concepts replace CUDA's thread/block/grid hierarchy. Refs are mutable memory references. because execution is sequential, each iteration safely accumulates without atomics. in CUDA you'd need atomics or a separate reduction pass. the memory model is also very different from NVIDIA's. zero hardware caches. VMEM is 32-128 MiB of software-managed scratchpad — 500-1000x larger than GPU shared memory per SM. all data must be explicitly DMA'd from HBM to VMEM before any computation touches it. four levels: HBM → VMEM → VREGs → MXU/VPU, plus SMEM for scalar control data. every byte of data movement is your responsibility. this is like CUDA shared memory except it's 500x bigger and there's no cache fallback. pipelining is mandatory. without double-buffering HBM→VMEM transfers, the MXU just stalls waiting for data. this is the single most important optimization on TPU. and because grid execution is sequential and deterministic, consecutive iterations that need the same input block skip the redundant HBM transfer automatically, impossible on GPU where block execution order is undefined. the compilation pipeline is unlike anything in this series: python → jaxpr → stableHLO → XLA HLO (71+ optimization passes) → LLO (78+ passes) → 322-bit VLIW bundles. the compiler packs instructions for scalar, vector, matrix, and DMA units into a single 322-bit word. everything in that bundle executes in parallel, with no runtime scheduling.show more

I told you to claim your free 16GB NVIDIA... GPU for learning Local LLMs. Now I’m going to show you how to double its inference speed without touching the hardware. Google Colab gives you an enterprise grade NVIDIA Tesla T4 GPU for free, roughly 4 hours every single day. It is the absolute perfect sandbox for learning AI engineering, testing inference flags, and pushing massive context windows. The local AI timeline is moving way too fast. If you aren't using Multi Token Prediction (MTP) yet, you are leaving massive performance on the table. I just pushed DeepMind’s Gemma 4 26B to 64.9 t/s on this exact free tier. Let's look at the raw benchmark data running on an Ubuntu Linux environment with the latest compiled llama.cpp binaries and quantized GGUFs from Unsloth via HuggingFace: # Qwen 3.5 9B (Dense): Base: [ Prompt: 626.7 t/s | Generation: 21.0 t/s ] With MTP: [ Prompt: 539.1 t/s | Generation: 24.8 t/s ] # Gemma 4 26B QAT (MoE): Base: [ Prompt: 634.2 t/s | Generation: 48.3 t/s ] With MTP: [ Prompt: 572.1 t/s | Generation: 64.9 t/s ] If you are paying attention, this single Colab notebook reveals 3 massive observations about the current state of local LLMs: # 1. The MTP Speedup (Software Overclocking) Standard autoregressive decoding guesses one token at a time. MTP acts like a highly optimized, built in speculative decoder. It predicts multiple future tokens at once and the main model verifies them in parallel. The result? Zero accuracy loss and a massive throughput increase. Gemma jumped from 48 to 65 t/s just by flipping a flag. # 2. The MoE Paradox (Bigger is Faster) How does a 26B parameter model absolutely destroy a 9B model in raw speed on the exact same hardware? Architecture. Qwen 3.5 9B is a dense model. it activates all 9 billion parameters for every single token. Gemma 4 26B is a Mixture of Experts (MoE) model. It routes data efficiently, activating only 4B parameters per token. You get the reasoning capabilities of a 26B model with the compute cost of a 4B model. 3. Thinking Efficiency When I ran the exact same complex prompt on both models, the larger MoE spent significantly fewer "thinking" tokens to arrive at the correct answer. A smarter model doesn't just give better answers; it gets to the point faster, saving you compute cycles and preserving your context window. # Want to run this yourself? Here are the exact llama.cpp CLI commands. For Qwen (MTP is baked into the main model): ./llama-cli -m Qwen3.5-9B-UD-Q4_K_XL.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 For Gemma (Using a separate lightweight draft model): ./llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --model-draft mtp-gemma-4-26B-A4B-it.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Stop waiting for a $3,000 rig. Boot up Colab, pull these models, and start building your stack. I’ve put together a completely free, cell by cell Google Colab notebook that automates this entire workflow so you can test it yourself in 5 minutes and learn. Link to the notebook is in the comments below. Experiemt with different MTP parameters, context windows and post your results in the comments.show more