Qwen3.8-27B Mac m4 max vs DGX spark: same engine (TensorFold 0.6.5 + DFlash2), same seeds: DGX Spark • 4-bit: 82.9 · 52 min (~69 tok/s) Mac Studio M4 Max • 4-bit: 83.6 · 71 min (~50 tok/s) • 8-bit: 84.7 · 95 min (~38 tok/s) Hardware moved it <1 pt. 8-bit bought +1.1 for 2x the memory. Spark-Bench v7, 80 tests, 0 errors. bf16 lands tomorrow. Run your own benchmarks:

Qwen3.8 Flash Next can nearly solve 50% of Terminal Bench 4.0, with 4 attemps, and probably even more with more attempts. Only 8 points below Astra's pass@1. That's extremely impressive for a model of that size, But it also helps explain why frontier models are already benchmaxxing this benchmark. Terminal Bench 4.0 isn't nearly as hard as it looks, especially when models get multiple attempts. This result requires xhigh reasoning effort. Medium is practically useless here, 4 medium attemps doesn't solve a single task that xhigh can't already solve in one attempt. More results and benchmarks here:

Qwen3.8-flash-next has a AA Intelligence index score of 40. For Qwen4-flash, I'm thinking 6 - 10 point increase, 46 - 50 range. Still not enough now that new Claude models are here, open models need to win back. Just barely beating Haiku-5.5 is not enough imo. Where do you think Qwen4-flash will land ? Will it beat Sonnet 5.5 ?

Qwen3.8-flash-next IQ3_S continues to amaze me. This is Qwen3.8-flash-next-GSQ-RSO-IQ3_S, running on my 3090 + 64GB RAM Strata backend of course. Goated One-shot, no external assets, three.js island, a ring-shaped coral atoll, turquoise lagoon, swaying palms, waves breaking on the reef, a sailboat drifting around The quality of waves, the swell timing and animation is really good, the palm trees getting this good was a surprise as well, super impressed by the result. Single file output, without using any coding or agent harness, this is raw result.

Qwen3.8 Flash Next thinking modes: Low is not always worse than Medium, but I couldn’t identify clear cases where Low should be preferred over Medium. So I’d recommend Medium as the default for cheaper reasoning, and Xhigh when higher accuracy matters, especially for coding. I could be wrong, but Low really looks like Medium with additional instructions (that you can find in the chat template) to reason less, rather than a separately post-trained reasoning mode.

Qwen3.8-Flash is 40% off through the rest of the month, making now a perfect time to run your evals. Designed for high-volume applications like coding and coworking assistants, @Alibaba_Qwen's model optimizes quality at a low cost. Try it today:

Qwen3.7-Max and Qwen3.8-Flash from @Alibaba_Qwen are now 40% off on serverless through September 30. Use Max for long-horizon agent work and Flash for high-volume, cost-sensitive workloads. Get started today!

Qwen3.8-2.4T on vLLM: a Pareto frontier spanning 5K total tokens/s/GPU at high throughput and 180 output tokens/s/user at low latency, across tuned PD configurations on @nvidia GB300 NVL72. Workload: 8K input / 1K output. Drawing on lessons from trial and error, we walk through the tuning decisions step by step: budget KV cache, benchmark prefill and decode separately, then choose topologies and MTP settings for each serving target. Deployment configs are included so you can reproduce the results. Great work from the @NVIDIAAI contributors and the vLLM community! Explore the frontier:

Qwen3.8-27B at up to 144 tok/s on an M5 Max MacBook Pro?! You have to try this Splash Engine! It's a new open-source inference engine called Splash. Instead of being a universal runtime like llama.cpp or Ollama, Splash optimizes the whole stack around the exact model 👇 ⚙️ model-specific kernels 🧠 hardware-aware memory planning 🚀 DFlash2 speculative decoding 💾 prompt-cache reuse 👥 continuous batching On the same 48GB M5 Pro running Qwen3.8-27B: 🚀 Splash: 74 tok/s ⚡ oMLX: 38 tok/s 🐌 Ollama: 24 tok/s At 32K context: 🚀 Splash: 54 tok/s And with 4 concurrent requests: 🔥 170 aggregate tok/s vs 43 tok/s for oMLX Now Inco + LM Studio are showing up to ~144 tok/s on M5 Max. And LM Studio Bionic 1.1.5 already added Splash as an experimental runtime. Requirements 🍎 M3 or newer 💾 36GB minimum 👍 48GB+ recommended Completely different performance because the software stack is optimized around what it is running. ⚠️ 144 tok/s is an Inco/LM Studio result. The detailed numbers above are from Inco's 48GB M5 Pro test. 🔗