Qwen3.8-Max Review 2026 — First Open-Weight Max-Class Model, #1 on the Agentic Index, $2/$6 Pricing
✅ Pros
- • First Qwen-Max-class model to go open-weight — the full Max weights drop next week, which is a first for the family and directly answers the community's biggest open-vs-closed question
- • Ranked #1 overall model by Artificial Analysis' Agentic Index, beating GPT-5.6 and Claude on agentic workloads — a meaningful signal for coding-agent users, not just chat benchmarks
- • Official reasoning_effort support (low/high/xhigh) lets you trade reasoning depth against cost on the same checkpoint, from quick classification to long-horizon autonomous work
- • Aggressive pricing at $2/M input and $6/M output with $0.25/M implicit caching — undercuts frontier Western APIs on agent-heavy token profiles
- • Qwen3.8-27B open weights ship next week too, giving local runners a path from the 27B sweet spot (huge on 4090s and 16GB Mac minis per HN) toward Max-class quality
⚠️ Cons
- • Open weights are announced but not yet released — 'next week' means you cannot self-host today, and HN is right to flag that the timeline can slip
- • The Max-Preview (July 19) → Max (Aug 6) transition was confusing: nothing on the official page documents what changed between the RL checkpoints, and HN users had to dig through Twitter to confirm it was a full release
- • Benchmark claims are hard to verify independently — Artificial Analysis' Agentic Index is credible, but Qwen's own page leans on self-reported numbers and the 'self-evolves through feedback loops' line drew justified HN skepticism
- • No dense mid-size option: the lineup jumps from 27B to Max, and multiple HN users (128GB RAM owners especially) said they'd want a 54B-72B dense model between the two
- • Agentic strength is concentrated in coding/cowork — the reasoning_effort tiers are a knob you must tune per task, and picking wrong costs you either accuracy or money
Teams and solo devs running agentic coding workloads who want frontier-class quality at Chinese-lab pricing — and anyone who has been waiting for an open-weight Max-class model to self-host or fine-tune
$2.00/M input tokens, $6.00/M output tokens, $0.25/M implicit cache reads via QwenCloud (qwencloud.com/models/qwen3.8-max); open weights free via Hugging Face when released next week
Quick Verdict
Qwen3.8-Max is the most capable model Alibaba’s Qwen team has shipped, and the first Max-class checkpoint to go open-weight. In the 48 hours after its August 6 announcement, it picked up 1,115 points on Hacker News and a second front-page story when Artificial Analysis ranked it #1 overall on its Agentic Index — ahead of GPT-5.6 and Claude in agentic workloads.
The two news items matter for different reasons. The launch post is about capability and openness: the full Max weights release next week, a first for the family, and Qwen3.8-27B goes open-weight in the same window. The Artificial Analysis ranking is about the shift everyone has been watching: Chinese labs’ open-weight models now lead the frontier on the exact workload — long-horizon agentic coding — that Western labs are betting their products on.
The price is the other headline: $2/M input, $6/M output, $0.25/M implicit caching on QwenCloud. For agent-heavy token profiles, that is aggressively cheaper than frontier Western APIs. 8.0: the announcement is 9, the shipping reality (weights not yet out, preview→release muddle) is 7.
What Qwen3.8-Max Actually Is
Qwen3.8-Max is the successor to the Qwen3.8-Max-Preview released July 19, and the flagship of the Qwen 3.8 family. The official line: it is the first Qwen-Max-class model to be released open-weight, with the weights arriving “next week” on open channels.
What’s new versus Preview. Qwen doesn’t document the delta explicitly, which HN noticed. The community reading: Preview models are earlier RL checkpoints; the official release is the checkpoint Alibaba is confident in after continued training. Practically, this is the version that ships in QwenCloud and powers the Qwen Code / Qwen Agent products.
The Agentic Index #1. Artificial Analysis’ Agentic Index measures end-to-end performance on agentic workloads — tool use, long-horizon task completion, coding-agent scenarios — rather than static chat benchmarks. Qwen3.8-Max taking the top slot is the strongest independent signal in the thread. One HN user put it plainly: “Qwen3.8-Max is the new flagship model for coding and harness systems… that’s gonna drop the price of LLM in the agent landscape a lot.”
reasoning_effort control. Qwen3.8-Max ships official support for reasoning_effort — xhigh (default, for complex tasks demanding thorough analysis), high, low — letting you tune reasoning depth per call to control cost. This matters for agent pricing: you can run classification and quick edits at low, and reserve xhigh for autonomous multi-step jobs.
Pricing & Plans
QwenCloud pricing for Qwen3.8-Max, confirmed via the model page and Alibaba’s own announcement:
| Item | Price |
|---|---|
| Input tokens | $2.00 / M |
| Output tokens | $6.00 / M |
| Implicit cache reads | $0.25 / M |
Context for that price: GPT-5.6-class and Claude-class frontier APIs sit at $3-15/M input and $15-60+/M output depending on tier. For an agent that burns hundreds of thousands of tokens per session, Qwen3.8-Max at $2/$6 cuts the API line item by a wide margin — which is exactly why HN’s thread keeps returning to “the price of LLMs in the agent landscape is about to drop.”
Self-hosted cost is not yet knowable: weights drop next week, so nobody has run benchmarks on local hardware yet. The 27B release in the same window is the immediate local play.
The HN Debate
The thread is a useful map of where open-weight adoption stands. Three arguments dominate:
1. Is the open-weights promise real? Yes — “the first time we will open-source the weights of a Qwen-Max-class model” is explicit in the announcement, and Qwen3.8-27B follows. The skepticism is about timing, not intent.
2. What changed since Preview? Unresolved. HN commenters compared the July 19 “Max-Preview” post with today’s “Max” post and found no public diff. One commenter noted the typical pattern — previews are earlier RL checkpoints, official release comes when the training run is finished — but the lack of documentation is a legitimate ding.
3. The local-model angle. Qwen3.6-27B is widely regarded as the best local model under 30B, and HN users are running it everywhere: IQ4 quantized on a 4090 with 24GB VRAM, the “bonsai” 1-bit quant fitting in 4GB on a 16GB Mac mini, and a 3.8-27B@fp8 on 96GB servers. The recurring ask: a dense 48-72B between 27B and Max. Alibaba didn’t deliver that, and users noticed.
The “self-evolves through feedback loops” line drew the sharpest pushback. HN’s response was essentially: every RL-trained model improves via exploration feedback; what matters is how long it runs without human intervention, not the marketing phrasing. Fair.
Who Should Use It
Get it via API now if you run agentic coding workloads — the agentic-index #1 ranking plus $2/$6 pricing is the strongest combination in the market today. Point your harness (Codex, Claude Code, Pi, whatever speaks OpenAI-compatible APIs) at QwenCloud and A/B it against your current model.
Wait one week if you want self-hosted: the open weights are the headline feature for local runners, and the 27B release gives the 4090/16GB crowd a concrete upgrade path.
Skip the hype on “self-evolution” — treat it as marketing; judge the model on the agentic index, the price, and your own eval runs.
FAQ
Is Qwen3.8-Max open source or just open weight? Open weight — the weights are Apache-style available for download and self-hosting, per the announcement, but this is not the full open-source stack (training data, code). The weights for Max and 27B release next week.
How does Qwen3.8-Max compare to GPT-5.6 and Claude? On Artificial Analysis’ Agentic Index it ranks #1 overall, ahead of both. On static benchmarks the picture is closer and more contested. Real answer: run your own evals on your actual agent workload before switching.
Can I run it locally? Not yet — weights ship next week. The 27B variant will be the practical local option (the Qwen3.6-27B lineage runs on 4090s and 16GB Macs at quantized sizes).
What is reasoning_effort? A per-call control for reasoning depth: low, high, or xhigh (default). Lower effort = faster and cheaper for simple tasks; xhigh for complex autonomous work. It is the same cost-tuning pattern frontier models use, now official on Qwen3.8-Max.
Where do I get it? API: qwencloud.com/models/qwen3.8-max ($2/$6 per M tokens). Open weights: next week via Hugging Face and Qwen’s channels.