The editorial argues the significance isn't raw parameter count but what it reveals about Alibaba's serving strategy: 95B active parameters means serving costs comparable to a dense 95B model while delivering knowledge capacity of something 25x larger. This doubles down on the DeepSeek economics playbook that broke pricing last year, pushing sparse MoE architecture harder with more experts and total capacity at the same activation budget.
The editorial highlights that Alibaba dropped a 2.4T-parameter model with no blog post, no benchmarks, no keynote — just a repo and shards. This pattern of shipping without marketing has become load-bearing for the ecosystem: the community learns what models can do from third-party evals rather than vendor copy, inverting the usual launch playbook.
The submitter simply posted the HuggingFace repo link with no editorializing, which itself validates the pattern — the community treats a bare Qwen repo drop as news-worthy on its own, driving 547 points without any accompanying announcement or narrative from Alibaba.
The editorial notes the HN thread split cleanly between people asking whether it runs on a Mac Studio (no) and those quietly downloading 4.8TB of shards to rented H200 clusters. Even sparse MoE requires holding all expert weights in fast-tier memory, meaning total parameter count still dictates hardware requirements regardless of the reduced active count.
Alibaba's Qwen team pushed `Qwen/Qwen3.8-2.4T-A95B` to HuggingFace this week with the usual amount of fanfare from that team: none. No blog post, no benchmark leaderboard screenshot, no keynote. Just a repo, a config.json, and shards.
The naming tells you everything the README doesn't. 2.4T total parameters, 95B active per forward pass — a Mixture-of-Experts architecture where the router picks a small subset of experts for each token. For comparison, DeepSeek-V3 shipped at 671B total with 37B active. Kimi K2 sits at 1T/32B. Llama 4 Maverick lands around 400B/17B. Qwen3.8 is roughly 3.6× the total parameter count of DeepSeek-V3 and 2.5× its active count — this is the largest open-weight model ever released, by a wide margin.
Within hours the HN thread had 547 points and the usual split: half the comments are people asking whether they can run it on a Mac Studio (no), and the other half are people quietly downloading 4.8TB of shards to a rented H200 cluster. The Qwen team's pattern of shipping first and writing the paper later has become a load-bearing part of the open-weights ecosystem — you learn what the model can do from evals other people run, not from marketing copy.
The interesting thing isn't the size. Anyone with enough GPUs can train something big. The interesting thing is what the size implies about Alibaba's serving strategy.
At 95B active parameters, Qwen3.8 costs roughly the same per token to serve as a dense 95B model — but has the knowledge capacity of something 25× larger. That's the entire pitch of sparse MoE, and it's the reason DeepSeek's economics broke the market last year. Qwen is now betting the same architecture, but pushed harder: more experts, more total capacity, same activation budget. If the router is good, you get GPT-4-class knowledge at Mistral-Small serving cost.
The catch is memory. Even sparse MoE has to hold all the expert weights in VRAM (or at least in fast-tier memory) because the router might call any of them for the next token. 2.4T parameters at fp8 is around 2.4TB of weights. That's 30 H100s minimum just to load, before you account for KV cache, activation memory, or throughput. Nobody outside a hyperscaler is running this locally. The Mac Studio crowd is out. The single-8xH100-node crowd is out. This is a model for cloud providers to serve, not for engineers to tinker with.
But that's fine, because the ecosystem has learned how to metabolize models like this. Within 72 hours you'll see: Together, DeepInfra, and Fireworks announce API endpoints; Unsloth ship dynamic quantization down to 2-3 bits; someone at ByteDance release a distilled 30B dense version that captures 80% of the reasoning quality; and llama.cpp merge a PR that lets you run a heavily-quantized version on 8× RTX 6000 Ada with 30 tok/s. The Qwen release cadence has trained the community to move fast.
What makes this release strategically pointed is the timing. OpenAI's GPT-5 has been out for two weeks and the reception has been "solid, expected, expensive." Anthropic's next Claude tier is rumored but unshipped. Google's Gemini 3 is API-only and priced accordingly. Into that gap Alibaba drops open weights of something that, on paper, has more capacity than any of them. The message to the frontier labs isn't subtle: the closed-weight moat is a race against a state-subsidized team that ships weights for free every eight weeks.
If you're building on Claude or GPT-4 today, nothing changes this week. Qwen3.8 isn't a drop-in — it's a bet you can make in six months when the tooling matures.
The practical playbook right now:
If you're a startup on an OpenAI/Anthropic bill above $50k/month, get an eval harness running against a hosted Qwen3.8 endpoint (Together, Fireworks, or DeepInfra will have it up by end of week) and compare it on your actual traffic. Not MMLU, not HumanEval — your actual traffic. The DeepSeek-V3 story last year was that people who ran the eval switched; people who read the benchmarks stayed. The gap between "benchmarks look competitive" and "we cut our inference bill by 60%" is almost always crossed by teams that run their own evals within a week of the release.
If you're doing local inference, this release is not for you, but the derivatives will be. Watch for a Qwen3.8-Distill in the 30-70B range in the next three weeks — that's the target size for a single 8xA100 node or a well-provisioned Mac Studio Ultra. The distill will land before the paper does.
If you're on the fine-tuning side, the interesting question is whether Alibaba shipped the base model or only the instruct-tuned variant. The HuggingFace repo naming suggests instruct-only, which limits what you can do with LoRA. If a base model follows in a few weeks, that's the real event — a 2.4T base model with permissive licensing would reset the open fine-tuning ecosystem the same way Llama 2 did in 2023.
If you're an infra engineer, the serving story is where the money is. vLLM's MoE support has matured a lot since DeepSeek-V3, but 2.4T parameters at fp8 across a 30+ GPU cluster with expert parallelism is not a solved problem. Expect posts in the next month from teams at Together and Fireworks explaining what they had to build. Those posts will be the actual documentation for how to serve this class of model.
The frontier-lab thesis has always been that scale is the moat and closed weights protect the scale investment. Qwen3.8 is the loudest counterargument yet: a Chinese team just shipped 2.4T parameters as a free download and didn't even write a blog post about it. If the evals hold up over the next two weeks — and Qwen's evals usually hold up — the pressure on OpenAI and Anthropic to justify their per-token pricing gets substantially worse. The moat isn't the weights anymore. It might not even be the training run. What's left is the product surface, the tool-use reliability, the latency, the enterprise contract. Which is a fine business to be in, but it's not the business those companies raised $10B to build.
Supposedly this is a Kimi k3 rival. Bit of a chonker, especially since they only released bf16 and fp8. So at launch this will be harder to serve than k3. No QAT on q4 means that someone with deep pockets (nvda?) will have to quant it, with plenty of calibration data. Should bring it ~1.3TB, so arou
Also of interest: DeepSeek V4-Pro-0813 (1.6T-A49B) benchmark scores have apparently just been announced on the DeepSeek WeChat channel and they're sitting about Fable 5 level.[1][1] https://www.reddit.com/r/LocalLLaMA/comments/1vmi0fg/deepsee...
https://unsloth.ai/docs/models/qwen3.8The 1bit quant model is at an astonishing 397GB with 95B active per MOE. This literally puts Opus 4.5 performance level into a machine a normal person could buy, and still gets usable tokens/second.The full lossless model BF16 is cl
> In particular, Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools, etc.That is unfortunate, that the open weight model doesn't have vision support or the 1
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Shall we bet on when the hardware needed for this (without quantizing and at good speed) will reach < 10k USD? I'm betting 2040. I can download it now, and then get the hardware later. Eventually we can all have these things running 24/7 in our home if we wanted to. I currently would no