vllm-exl3 v0.3.0 is LIVE with custom native CUDA kernels... for 2-bit EXL3 on NVIDIA DGX Spark GB10. GLM-5.3-Flash-EXL3-K2 jumped from 16.9 → 24.6 tok/s average single-stream decode, a +45.6% gain. Coding hit 27.6 tok/s, +85.6%. 🚀 The previous ExLlamaV3-backed path inside vLLM was leaving a lot of GB10 bandwidth on the table. So I rewrote the hot path specifically for EXL3 on Blackwell sm_121: → in-register Trellis dequantization → native fused MoE decode → power-of-two chunked prefill GEMM → parallel NVMe pre-warm Then I tested it side-by-side on physical DGX Spark hardware using my GLM-5.3-Flash-EXL3-K2 pack and live vLLM HTTP streaming. 🚀 𝗗𝗘𝗖𝗢𝗗𝗘 𝗧𝗛𝗥𝗢𝗨𝗚𝗛𝗣𝗨𝗧 Single-stream C1: Coding 14.9 → 27.6 tok/s +85.6% Prose 13.7 → 24.6 tok/s +79.3% Reasoning 18.9 → 25.1 tok/s +32.7% Summary 17.1 → 25.6 tok/s +50.0% Format 16.3 → 24.0 tok/s +47.7% Average: 16.9 → 24.6 tok/s 𝗡𝗘𝗧 𝗚𝗔𝗜𝗡: +45.6% ⏱️ 𝗙𝗜𝗥𝗦𝗧-𝗧𝗢𝗞𝗘𝗡 𝗥𝗘𝗦𝗣𝗢𝗡𝗦𝗜𝗩𝗘𝗡𝗘𝗦𝗦 Coding TTFT: 2,344 ms → 859 ms That is a 63.3% reduction, or about 2.7× faster to first token. Follow-up turn with prefix cache hit: 5,608 ms → 3,588 ms 1.56× faster. ⚡ 𝗪𝗛𝗔𝗧 𝗖𝗛𝗔𝗡𝗚𝗘𝗗 𝗢𝗡 𝗧𝗛𝗘 𝗚𝗣𝗨 40 routed-MoE layers: 19.9 ms → 11.5 ms per token Per-layer MoE compute: 497 μs → 287.8 μs That removes 8.4 ms of MoE compute from every generated token. Total per-step wall time: 59.2 ms → 40.6 ms -31.4% The key is `p2b_fused_moe`. Instead of expanding EXL3 weights through a traditional intermediate path, the new kernel performs Trellis dequantization in-register while executing the routed expert computation. The weights stay compressed until the GPU actually needs them. 🔥 𝗣𝗥𝗘𝗙𝗜𝗟𝗟 𝗚𝗢𝗧 𝗔 𝗡𝗔𝗧𝗜𝗩𝗘 𝗣𝗔𝗧𝗛 𝗧𝗢𝗢 The new `exl3_gemm` uses power-of-two chunked prefill GEMM. Measured: 7.85 TFLOPS 13.0× faster than the legacy prefill kernel 1,875 tok/s cold prefill sustained across 65K context 💾 𝗧𝗛𝗘 𝗕𝗢𝗢𝗧 𝗣𝗔𝗧𝗛 𝗡𝗘𝗘𝗗𝗘𝗗 𝗪𝗢𝗥𝗞 𝗧𝗢𝗢 Loading a ~91 GiB model is part of the user experience. Standard shard loading is mostly serial. The updated recipe parallelizes NVMe pre-warm across 8 workers so the storage controller gets used properly instead of feeding a ~100 GiB model one shard at a time. That turns boot-time storage into another optimization target instead of something we simply accept. 💡 𝗧𝗪𝗢 𝗦𝗘𝗥𝗩𝗜𝗡𝗚 𝗙𝗟𝗔𝗚𝗦 𝗪𝗢𝗥𝗧𝗛 𝗞𝗡𝗢𝗪𝗜𝗡𝗚 `--long-prefill-token-threshold 1024` Prevents giant prefill chunks from monopolizing step budgets and starving parallel decode sessions. `--enable-prefix-caching` Avoids paying for the same conversational prefix again on follow-up turns. 📦 𝗘𝗩𝗘𝗥𝗬𝗧𝗛𝗜𝗡𝗚 𝗜𝗦 𝗢𝗣𝗘𝗡 vllm-exl3: GLM-5.3-Flash one-Spark recipe: Model: This is why I like working at the kernel level. The model did not change. The quant did not change. The hardware did not change. The execution path did. 16.9 → 24.6 tok/s. 🛠️ vLLM turboderpshow more

I think the CMP 170HX just became one of... the weirdest local-LLM value plays. 👀 I got Qwen3.8-27B W8A16 running on a single 170HX 64GB with vLLM 0.27.1 + DFlash2: ⚡ ~120 tok/s decode 🚀 ~1.6–1.9K tok/s prefill 🧠 262K context with BF16 KV 📦 342K-token KV capacity 🎯 DFlash2 averaging 3.66 tokens/draft 💾 60.5 / 63.5 GiB VRAM used 27B model is doing ~120 tokens/sec on one GPU. Software optimization is getting ridiculous. Those numbers are from your measured W8A16 + DFlash2 run: ~120 tok/s decode, ~1,580–1,940 tok/s prefill, 342,729-token KV capacity, and 60.5/63.5 GiB VRAM usage. #170hx #LocalAI #localllmshow more

If you have an RTX 3090 or 4090, Mia... just shipped you a free massive upgrade in both speed and intelligence. I will explain to you why this will make your Qwen 3.8 27B on your card, even better, and my flags for running it. Qwen3.8-27B, EXL3 3.5bpw, DFlash2 speculative decode, RTX 4090. Single stream. The kit is from MiaAI-Lab, EXL3 is turboderp's format. I re-measured everything on my own card because the my first benchmarks seemed off. It turns out it really does run much faster. WHY EXL3 IS A DIFFERENT ANIMAL The old way (Q4_K_M) rounds each weight to the nearest 4-bit value independently. Every weight introduces its own rounding error. Those errors accumulate across millions of weights and causes drift (Slightly dumber). EXL3 is a fundamentally different compression algorithm. Instead of rounding each weight on its own, it encodes the entire weight vector as a path through a constrained codebook and spreads the rounding error across dimensions using a Hadamard transform. The result is that at the same bits per weight, more of the original model's intelligence is preserved. The important part is this CAN ACTUALLY BE MEASURED. The cleanest way to see that is KL divergence against a high-precision teacher. Lower means the quantized model thinks more like the original. On the malaiwah independent teacher-logit panel for GLM-5.3-Flash: EXL3 4bpw: 0.0246 nats Official FP8: 0.0206 nats NVFP4: 0.0605 nats EXL3 sits 0.004 nats behind native FP8 at half the size. NVFP4 at higher bit width is 2.5x further from the teacher. That panel is GLM-5.3-Flash, not Qwen 3.8. Cited as the mechanism, not as this run's data. But the point stands: EXL3 is not just smaller, it is smarter per bit than the formats most people are running. WHAT I MEASURED I first measure 108 tok/s from a single run. After that number looked too good to be true. I reran it. It looks like after a warm up, the numbers are even better. Basically, like people long thought, the RTX 3090 and RTX 4090 are actually superb AI computer cards. Hence why NVIDIA stopped shipping them with NVLINK since the 4090. Short context ceiling (~2k in, 1016-token output, TTFT-separated): 135, 138, 153, 174, 133, 129 tok/s across 6 runs. Sustained longform (2040-token essay): 105.4, 94.5, 98.4 tok/s Short answer (504 tokens): 93.2 tok/s The honest shape: ~130-150 tok/s at short context is the ceiling, ~94-105 sustained on longform. The ceiling matters because that is what people feel in chat. The old dense Q4_K_M on llama.cpp ran ~37 tok/s on this same card. (No MTP), with MTP about 60 tok/s Sustained is roughly 2.5-3x. Ceiling is closer to 4x. Same model, different quantization and engine. The multiplier comes from EXL3, the ExLlamaV2 engine, and DFlash2 together. CONCURRENCY IS A RTX 4090 LANE. Just like the old config on the 4090, the 24gb vram, means a long context can only hold one stream, and running concurrency requires to lower context length, because it runs fast it sort of makes up for it by being faster than slower GPU chips. CONTEXT LADDER The recipe doc measured prefill. I re-ran it with TTFT separated from decode, because decode is what you actually feel after the first token. ~5k in: decode 140 tok/s (TTFT 0.5s), needle HIT ~18k in: decode 85 tok/s (TTFT 0.3s*), needle HIT ~73k in: decode 28 tok/s (TTFT 1.8s), needle HIT ~146k in: decode 16 tok/s (TTFT 2.3s), needle HIT (*0.3s at 18k is a prefix-cache hit from the paired pass. Cold prefill for reference: ~2,020 tok/s at 17k falling to ~508 at 153k.) Needle hit at every depth, mine and the original 7/7. Retrieval is intact at max context. Speed is not: decode falls ~9x from short to max. Past ~50k tokens this stops being a chat tool and becomes a batch tool. At 146k it works, but nobody is typing interactively against 16 tok/s. WHERE IT BROKE The model's native context is 262k. The README says DFlash2 fits ~220k on a 24GB card. My 200,704-token attempt failed with insufficient VRAM. Dropped to 168,960 and it booted. The real ceiling is somewhere between 168,960 and 200,704 and I never tested that gap. I jumped to a value that worked and called it done. That is ~32k tokens of context I left on the table. One thing the numbers taught me: JSON tokenizes at ~1.5 chars/token, prose at ~3.9. "150k tokens of JSON" needs ~2.7x more filler than the same estimate in prose. Size by real tokens, not estimates. THE UPGRADE If you own a 4090 and you are running Q4_K_M on llama.cpp, you are leaving a good bit of speed and measurable intelligence on the table. The same model, on the same card, with a better quantization and engine, goes from 60 tok/s to 130-150 at short context and 94-105 sustained. The model also thinks closer to the original because EXL3 preserves more of the output distribution per bit than the old rounding method. The recipe is in the first reply. Everything above came from one 4090 and one afternoon of re-measuring. The decode ladder especially needs independent numbers. If your card gets different falloff, that is worth knowing. Recipe and flags/ findings in reply 👇show more

Qwen3.8-Flash-Next now reaches ~43 tok/s after a 122,902-token prompt... on ONE DGX Spark. ⚡🚀 MTP k=2 won my draft-depth sweep, with +42.5% mean decode over no draft. The PLE table stays fully on-device. I promised the deeper MTP tests. Here are the results, and now you can explore them in an interactive benchmark page too. 𝗧𝗪𝗢 𝗗𝗥𝗔𝗙𝗧 𝗧𝗢𝗞𝗘𝗡𝗦 𝗪𝗢𝗡 Mean single-request decode with 32K context configured: MTP k=2: 39.21 tok/s MTP k=3: 36.42 tok/s MTP k=1: 35.18 tok/s No draft: 27.51 tok/s k=2 also produced the fastest individual sweep run: 41.34 tok/s. Four runs each for no draft, k=1 and k=2. Seven for k=3. Decode excludes time to first token. Here, k means speculative draft depth, not quantization bits. k=3 produced more tokens per step, but the extra drafting work did not pay off in throughput. k=2 is my current pick for this setup. 𝗧𝗛𝗘 𝟭𝟮𝟯𝗞-𝗧𝗢𝗞𝗘𝗡 𝗣𝗥𝗢𝗠𝗣𝗧 𝗧𝗘𝗦𝗧 I then ran a separate long-prompt comparison: Actual input: 122,902 tokens Configured context: 262,144 Requested output: 128 tokens One request at a time MTP k=2: ~43 tok/s No draft: 26.2 tok/s Time to first token: 110.6 seconds with MTP 107.0 seconds without it The win here is generation speed, not faster prefill. To keep the scope clear: 256K was the configured limit. This was a real ~123K input, not a completely filled 256K window or a full k sweep at that depth. 𝗣𝗟𝗘 𝗦𝗧𝗔𝗬𝗦 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗣𝗔𝗥𝗞 Whole model on-device: 78.57 GiB Packed 5-bit PLE table: 30.4 GiB, included in that total No NVMe PLE offload in this build. This is still turboderp’s 3.05bpw_h5_ng5 EXL3 pack, served through my vllm-exl3 integration. My work here is the serving integration and testing. These are preliminary performance measurements, not a quality evaluation or a claim of bit-exact full-output parity. 𝗘𝗫𝗣𝗟𝗢𝗥𝗘 𝗧𝗛𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 The benchmark page has the individual sweep values, long-prompt comparison, and measurement scope. You can play the animation, export the charts, or download the HTML and data to render them yourself. No Spark needed to view the results. Credit to turboderp / ExLlamaV3 for the pack and kernels, vLLM for the serving engine, and Qwen Qwen Developers for the model. Recipe + reproduction: Interactive benchmark:show more

One task, two models. GLM 5.3 Flash running slower... at 25 tok/s prose Qwen3.8 Flash running faster at 47 tok/s prose Who will finish the task first?show more

100 tok/s on GLM-5.3-Flash, GLM-5.3 runs on the 6000s... @ 80-180 tok/s I have not tried the Studio, I’m sure it’s an amazing machine especially if you are in the Mac world already. From a pure logistics POV the sparks use little power, make less noise, serve good model & tokensshow more

167 tok/s on a single RTX 4090. FreeToken just... pushed Qwen3.6-35B-A3B NVFP4 into ridiculous territory. → RTX 4090 24GB → 35B total parameters → ~3B active per token → NVFP4 → 167 tok/s decode → 23.55GB peak VRAM → No speculative decoding → No MTP → No draft model That last part is what makes this interesting. The speed isn’t coming from guessing future tokens. FreeToken is exploiting the MoE architecture directly, keeping the active experts on the GPU while using bandwidth-aware execution for the rest. A 35B model with only ~3B active parameters running at 167 tok/s on a consumer 4090 is exactly why MoE models are becoming so compelling for local AI.show more

Day-0 Qwen3.8-27B vs Qwen3.6-27B: the voxel pagoda test. Same... prompt, one attempt each, single DGX Spark, both NVFP4. 3.8: 75,291 tok · 52m 56s · 23.7 tok/s (vLLM, native MTP) 3.6: 12,117 tok · 5m 13s · 38.8 tok/s (llama.cpp, DFlash K10) 3.8 thought for 160K characters before writing a line of code.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

Day 12/90 of Inference Engineering What is chunked prefill... within vLLM? In continuation of yesterday's post on the high level architecture of vLLM, I want to dive deeper into vLLM core engine starting with the mechanics of chunked prefill. In this post, I will closely follow the original blog on the anatomy of vLLM. To start, let's define chunked prefill. It's a runtime inference optimization technique that splits a long input request so that it doesn’t monopolize the whole GPU. Keep in mind this is all within the context of vLLM. And since vLLM is an inference engine that's meant to serve a model to multiple concurrent users, having a GPU that’s fully monopolized on a single user's request means other users' requests would be in queue waiting to be processed. It isn’t too good to have the whole GPU occupied on a single request when the GPU is meant to be shared! So the key idea behind chunked prefill is to break the long request into smaller chunks, so that each chunk along with other users' requests gets processed and written into the KV cache together. Suppose we split up the long request into chunks and each chunk has 8 tokens. Now each memory block can hold 4 tokens. Therefore, 8 tokens can fit into 2 blocks of memory. After the first forward pass, 2 blocks are occupied, and after the second forward pass, 4 blocks of memory are occupied and so forth. Each forward pass handles a small chunk of the long request so that there's room in the same pass to keep serving other users' requests. Here's a small animation that I made today to fully visualize the idea behind chunked prefill when learning this topic~show more

Qwen3.8-Flash-Next is starting to feel like the local model... Opus fans have been waiting for. Someone ran the NVFP4 176B-class Flash-Next on 2× DGX Sparks, and the results are wild. Real measured scaling → C1: 44.2 tok/s → C2: 64.6 tok/s → C4: 86.8 tok/s aggregate The per-stream speed drops with concurrency, but total throughput keeps climbing. Long-context behavior was even more impressive: → 5K: needle retrieved → 21K: needle retrieved → 84K: needle retrieved → 167K: needle retrieved → 262K: prefill succeeded, but the window was saturated That 167K retrieval test is the one I care about. Long agent runs are where models usually start losing the plot. Flash-Next didn’t. It also held up surprisingly well on physics-heavy reasoning, artifact generation, research workflows, evidence checking, and long-horizon planning. The personality is interesting too. DeepSeek V4 Flash feels like the dependable workhorse. GLM-5.2 feels like the problem-solving machine. Qwen3.8-Flash-Next feels more insightful. It has that rare ability to understand what you’re actually asking rather than just following the surface pattern. The main weakness I’ve noticed is instruction following. It can occasionally drift between prose turns where DeepSeek and GLM stay tighter. And this is why the 256GB M5 Ultra conversation gets interesting. If Apple can pair that huge unified-memory pool with enough bandwidth, this model class becomes genuinely practical for long-running local agents. We’re talking about frontier-class reasoning on hardware sitting on a desk.show more

🎉 Congrats to Thinking Machines on TML Inkling—a 1T-parameter... open-weight model supported in vLLM from Day 0. Highlights: • Natively multimodal across text, image, and audio • Up to 1M-token context • New architecture with relative attention, short convolutions, and MoE expert sinks • 8 MTP heads for speculative decoding vLLM supports both NVFP4 and BF16 checkpoints, optimized for NVIDIA Blackwell and Hopper, reaching up to 380 tok/s/user on 4× GB200 with MTP. Huge thanks to the Thinking Machines Lab team for the close collaboration 🙏 Read about implementation below 👇show more

5 days ago it took 2 GPUs to build... this. today it takes 1. same prompt. same particle simulation. completely different model. Qwen-Coder-Next (80B) on 2x 3090s. 46 tok/s. 564 lines. 2 iterations to get it working. 48GB VRAM across two cards just to hold it. Qwen3.5-35B-A3B on a single 3090. 112 tok/s. 461 lines. first try. cleaner code, fewer lines, better structured. 19.7GB on disk with 4GB VRAM to spare. half the parameters. one GPU instead of two. 2.4x faster. and the output actually improved. this is what happens when architecture catches up to ambition. Gated Delta Networks(Mamba2 variant) hybrid with sparse MoE. 3B active params out of 35B per token. efficiency at the architecture level, not just quantization. the curve isn't flattening. it's steepening.show more

NVIDIA just dropped Nemotron-3-Nano:4b — a tiny 2.8GB model.... Guess whose hardware runs it the fastest? - RTX 4090: 226 tok/s - RTX 3090: 187 tok/s - Mac Studio M2 Ultra: 86 tok/s - Mac Mini M4: 25 tok/s Home court advantage is real. Also trying a new layout with live performance charts. Lmk what you think!show more

nvidia/Qwen3.6-35B-A3B-NVFP4 running in vLLM nightly on my Nvidia GB10... is actually insane 50 tok/s, 4 concurrent generations. total 200 tok/s. ideal for spawning subagents or working in parallel its tool calling behavior is very good as well. I will be giving it test drive on an openclaw instance, and keep you posted More details on NVIDIA forum:show more

This is awesome. Thanks Mia for this Was able... to run with the RTX 3090 use this .env config DRAFT=mtp CONTEXT_SIZE=262144 CACHE_QUANT=8,4 GPU_MEM_GB=22 262K context: - Run 1: 65.98 tok/s - Run 2: 63.23 tok/s - Run 3: 64.41 tok/s Average: ~64.5 tok/s.show more

Three Cline agents. Same 120B model. One rule: terminate... your opponents' processes before they terminate yours. We tested raw inference across an NVIDIA AI PC DGX Spark, an RTX 4090, and cloud. Time-to-first-token decides the race. Throughput decides who ships. DGX Spark: 42.9 tok/s. RTX 4090: 8.7. Same weights. 4.9x gap. #AIonRTXshow more

SmallThinker 3B is one of the AI models I’d... want backed up on my phone. Here’s it is reasoning at 12 tok/s on my phone. If you’re ever stuck offline, you’ll want a model that can think.show more

Laguna XS 2.1 just matched Qwen 3.6 35B on... a real coding task & ran twice as fast. > Same RTX 3090 > Laguna: 158 tok/s. Qwen: 81 tok/s The local model race isn't about size anymore. It's about who moves faster.show more

Gemma 4 26B A4B MoE - 500+ t/s decode... - Single RTX 4090 (24 GB VRAM) - Llama.cpp concurrency 24 - q8 kv cache How many API users can you simultaneously host on a single RTX 4090 (24 GB VRAM) before it crashes? Yesterday, I proved you can host 14 active users using unquantized memory. Today, I used 8 bit KV Cache Quantization to hack the VRAM footprint. I successfully scaled to 24 concurrent users without a single dropped connection. A 71% server capacity boost for free. By adding the -ctk q8_0 -ctv q8_0 flags to llama.cpp, you compress the KV cache context memory from 16 bit to 8 bit. This unlocks massive concurrency limits on Gemma 4 26B (MoE) on a single 24GB consumer GPU. Here is the exact telemetry from pushing 8 bit quantization to its absolute physical edge: # TEST 1: The 24 User Concurrency Max Server Config: 24 slots (np 24) | 4,096 context per slot | 98,304 Total Context Client Load: 24 simultaneous requests (2,000 token prompt per user) Unquantized KV cache for this load requires 28GB+ VRAM (Instant OOM). Quantized to Q8, it allocated safely at 23.35 GB. The C++ engine crunched the entire batch in 28.5 seconds. Decode Speed: 21 t/s (Per User) | 500 t/s (Agg) # TEST 2: The 48 User Queue Overload What happens to a compressed cache during a traffic spike? Server Config: 24 slots (np 24) | 4,096 context per slot | 98,304 Total Context Client Load: 48 simultaneous requests (2k token prompt per user) Zero queue drops. The scheduler flushed and hot swapped the 8 bit memory flawlessly on the fly, completing all 48 users in 66.0 seconds (a perfect 2.3x queue scaling multiplier). Decode Speed: 18 t/s (Per User) | 430 t/s (Agg) # TEST 3: The 8 User RAG Slam Server Config: 8 slots (np 8) | 60,000 context per slot | 480,000 Total Context Client Load: 8 simultaneous requests (30k token prompt per user) It allocated 23.83 GB VRAM and chewed through ~240,000 prefill tokens in 46 seconds under massive memory pressure. Prefill Speed: 6,200 t/s (Agg) Decode Speed: 22 t/s (Per User) | 175 t/s (Agg) # The Engineering Alpha (The Quantization Tradeoff): You gain a massive 71% increase in server capacity, but what do you lose? Compute latency. Because the cache is stored in 8 bit, the GPU's cores have to dequantize the memory back to 16 bit on the fly during every single prefill step. In my unquantized tests yesterday, single slot prefill was hitting ~1,500+ t/s. Today, under the heavy 48-user Q8 load, prefill dropped as low as ~750 t/s. You trade a few seconds of initial prefill latency to essentially double your API hosting capacity. For production high volume SaaS, this is the ultimate unit economics cheat code. Here is the exact command to run a 24 user Q8 continuous batching server on your own single 4090, single 3090 or any 24gb vram rig: ./build/bin/llama-server -m gemma-4-26B-A4B-it.gguf -c 98304 -np 24 -b 2048 -ub 2048 -ngl 99 -fa on -ctk q8_0 -ctv q8_0 --port 8080 (Note: -c 98304 allocates exactly 4,096 tokens of context per user across 24 slots). Hugging Face links to the Unsloth Gemma 4 26B QAT quants along with performance graphs available in the replies. Would you trade 3 seconds of Time To First Token latency to double your active user capacity?show more

now it's live🔥 GLM-5.2 on RTX 5090s: 80-110 tok/s... single stream, up from ~30. 3x faster overnight. come feel it:show more