Qwen4-Max released today: 60% cheaper per task
Qwen4-Max launched today: 60% cheaper per task, but what I'm more excited about is the locally deployable 27B. This morning I came across the Qwen4-Max launch, and the first thing I looked at was the billing section—the spec sheet came second. Alibaba has brought its new-generation foundation model to the table this time, and the official blog title reads, "Your new AI colleague has arrived." 2.4 trillion total parameters sounds intimidating, but each token only...
Qwen4-Max Launches Today: 60% Cheaper Per Task, But What I'm Really Waiting For Is the Locally Deployable 27B
When I saw the Qwen4-Max launch this morning, the first thing I flipped to was the pricing section — the spec sheet came second.
Alibaba has brought out its new-generation foundation model, and the official blog headline reads: Your new AI colleague has arrived. 2.4 trillion total parameters sounds intimidating, but only 71 billion are activated per token — down from roughly 95 billion in the previous generation Qwen3.8-Max. Waking up a quarter fewer parameters, yet doing the job better. It hit sixth place on CodeArena, the global blind front-end coding benchmark, wedged between GPT6-Astra and Claude Fable5.1 — the best ranking any Chinese model has achieved on that third-party leaderboard. It natively handles images, with a context window of 1 million tokens in one go. The examples in the official blog are pretty concrete too: Blender modeling, Unity scene building, one-sentence generation of playable 3D games, and long-horizon Agent tasks that are far more stable than the previous generation.
I won't dig too deep into the architecture. The official description is GDN plus Qwen sparse attention, four parallel residual branches, and a 328B N-gram Embedding — essentially overhauling sparse MoE's attention, residual, and embedding all at once, so this 2.4T behemoth only wakes up the parts it needs to. What I care about more is the side effect: it talks less.
There's another set of numbers in the official blog that I quite like. They built a framework called Work360 that breaks real work into 26 top-level domains, 182 second-level categories, and 721 sub-directions. Qwen4-Max scores 52.69 on this evaluation, up from 47.74 for the previous generation. On concrete tasks it's even more intuitive: in automotive manufacturing, machining programming for a single workpiece takes a human 475 minutes, but the model compresses it to 24 minutes; in financial research, it can call 90 MCP tools and directly spit out Excel workpapers and PPTs. These are official self-reported benchmarks, though — how much you actually save on your own work, you'll have to run it yourself to find out.
What's the biggest fear when running Agents? The model rambles on and on, tokens burn like crazy, and the money's gone before the job is done. According to the official blog, on the Terminal-Bench 4.0 command-line test, with the same Xhigh thinking level, average token usage per attempt dropped from 289K to 151K — nearly halved — while the score rose from 37.9 to 48.5. The official word is roughly 48% and 49% reductions at the two reasoning intensity levels respectively. Talking less while scoring more — that I genuinely didn't expect.
Prices came down too. On the domestic Qwen AI platform, it's 10 yuan per million input tokens and 30 yuan for output, down from 12 and 36 for the previous generation; implicit cache hits dropped from 1.5 yuan to 0.5 yuan — more than a 70% discount. On the international site, it's $1.5 input, $4.5 output, and $0.075 for cache. Fewer tokens consumed plus lower prices: in the official Work360 evaluation, the average cost of completing a complex task fell from $4.43 to $1.67, a 62% drop. Domestic media converted it to 26.58 yuan down to 11.16 yuan, a 58% drop — same story, just different accounting.
In the past, running long-cycle workflows always required careful budgeting. Now that per-task costs have more than halved, CLI and Agent workflows have a whole new tier of possibilities to play with.
It's not hard to get started now either. I clicked through all the entry points: QwenCloud's API and Token Plan, QwenWork for office use, and Qoder are all open, and the Qwen App and Qwen Studio are ready to use directly. For the open-source crowd, the key line came at the end of the official post: small and medium-sized models will be open-sourced progressively, sharing the same architecture and training methodology as Max. Max itself won't be open-sourced for now.
But I suspect most local players, like me, after reading through the press release, have their minds on something else: the locally deployable Qwen4-27B.
Let me be precise here. What the local community currently treats as the top pick is the previous generation Qwen3.8-27B, which has built up a solid reputation for both performance and deployment friendliness. Qwen4-27B and Flash still hadn't been released as of this morning; the official word is only that small sizes will be open-sourced progressively — no mention of 27B specifically, no date given. The "about four days after Max" figure circulating online is community convention extrapolated from previous generations' cadence, not an official announcement. I'd advise against marking your calendar too firmly.
Why the wait? The 27B size hits a real sweet spot. Don't be scared off by the 24GB VRAM figure — based on the community's estimates for quantization tiers, a machine with 16GB of RAM can run it. Ordinary laptops and mini-PCs can reach it; no need to stack multiple GPUs or rent cloud instances. Long context, tool calling, local knowledge bases — all of it can run behind closed doors. The previous generation proved this path works; as long as the new generation inherits the "burn fewer tokens" capability, the cost of running long-cycle Agents locally drops to nearly zero. My bet is 27B will be the breakout star of this wave.
Max is usable right now — hand the long tasks it can handle to the API, and wait for the small size to come home for the ones it can't. Neither side holds you up. What I'm more curious about is whether, when 27B actually goes open-source, a 16GB machine can swallow the 1M context too. I don't have an answer for that yet — once the weights land, I'll run it first thing and show you.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.