Qwen3.8-Flash Costs a Tenth of Qwen Max, Once You Fix the FP8 Bug

Quick verdict

Alibaba dropped Qwen3.8-Flash into OpenCode Go, and the number that matters is the price. It runs at roughly $0.15 per million input tokens and $0.47 per million output, which independent testers put at about 20 times cheaper and twice as fast as Qwen3.8-Max. The model is a 125B mixture of experts with only 6B active, a 1M-token context, and multimodal input. The catch showed up fast: at FP8 the model lost track of multi-turn conversations until users switched the KV cache away from turboquant to BF16. Once they did, it held together. So this is a genuinely cheap long-context model with a config trap you need to know about before you route real work to it.

What actually shipped

One model, aimed squarely at the cheap tier. Qwen3.8-Flash sits under the Max line as the high-volume option, and the design choices all point at cost per token rather than a top leaderboard spot.

• 125B total parameters with 6B active. That small active count is what makes it fast and cheap to serve, since each token only lights up a fraction of the model.
• A 1M-token context window and multimodal input, so it takes the same long documents and images the Max tier does.
• Pricing that skalskip92 clocked at about $0.15 per million input tokens and $0.47 per million output, roughly a twentieth of what Qwen3.8-Max charges at $2 in and $6 out.
• Throughput around 2x faster than Max in the same tests, which tracks with the tiny active-parameter count.
• Shipped straight into OpenCode Go, so it landed in an agent harness rather than as a bare checkpoint.

The early field reports were not uniformly good, and that is the part worth reading closely. QuixiAI reported that multi-turn tracking broke at FP8, the format most people reach for to serve a model cheaply. The fix turned out to be the KV cache: switching it from turboquant to BF16 cleared the problems, and the broader recommendation that followed was to run BF16 KV plus optional CPU offload for stability. In plain terms, the cheapest way to serve it is also the way that breaks it, so the real cost sits a little above the headline number.

Why it matters

A tenth to a twentieth of the price changes what you can afford to automate. Work you would never send to a $6-per-million model, like summarizing long threads, triaging tickets, or first-pass code review across a whole repo, becomes cheap enough to run on every item instead of a sample. That is the same shift the open Chinese labs keep forcing, and it is why the paid frontier keeps having to justify its markup. We walked through the last round of that pressure in DeepSeek V4-Flash shipping open weights.

The FP8 caveat matters more than it looks. A cheap model that quietly loses the thread halfway through an agent run is worse than an expensive one that does not, because you pay for the failed turns twice, once in tokens and once in cleanup. The lesson is the same one that keeps coming up with fast open models: the serving config is part of the product, not a footnote. If you are weighing which tasks belong on a cheap tier and which do not, the routing case is in cutting AI coding agent costs with model routing.

The wider pattern is that the cheap tier is filling in fast. A 1M-context multimodal model at 15 cents a million would have been a headline on its own a year ago. Now it shows up as the budget sibling of a launch that already happened, on a quiet day. We track where that leaves the open ecosystem in open-source AI models in 2026, and if you would rather test several of these without committing to one lab, running them behind one subscription is covered in the best app for running multiple AI models.

Video: cheap open models and the flash tier

This covers where the cheap, long-context open models sit against the paid frontier and how to think about routing work to them.

FAQ

How much cheaper is Qwen3.8-Flash than Qwen3.8-Max?

About 20 times cheaper on the tests that have surfaced so far. Flash runs around $0.15 per million input tokens and $0.47 per million output, against $2 and $6 for Max. It is also roughly twice as fast, which comes from having only 6B active parameters out of 125B.

What is the FP8 problem people are hitting?

At FP8, users reported the model losing track of multi-turn conversations. The fix was to switch the KV cache from turboquant to BF16, which cleared it up. The practical advice from testers is BF16 KV cache plus optional CPU offload, so budget for slightly more than the cheapest possible serving setup.

Is it good enough to replace a paid coding model?

For high-volume, forgiving work it is worth trying, since it shipped straight into OpenCode Go and the cost of testing is low. For anything where a dropped turn is expensive, keep it on the tasks that fit and save the paid tier for the rest. That split is the point of the best AI coding agents guide.

How does it compare to DeepSeek and GLM's cheap tiers?

It is the same design point as GLM-5.3-Flash and DeepSeek's flash models: small active-parameter MoEs with long context and aggressive pricing. Qwen leans on the 1M context and multimodality here. For the GLM side of that race, see GLM-5.3-Flash at nine cents a task.

Sources

• @Alibaba_Qwen - Qwen3.8-Flash in OpenCode Go, 125B/6B, 1M context, multimodal
• @skalskip92 - roughly 20x cheaper, 2x faster, ~$0.15/$0.47 per million
• @QuixiAI - broken multi-turn tracking at FP8
• @QuixiAI - BF16 KV cache fixes it, recommends BF16 KV plus CPU offload

Further reading

• Qwen3.8-Max going open, and the 27B that fits in 17GB
• DeepSeek V4-Flash shipped open weights the same day
• Cutting AI coding agent costs with model routing
• Open-source AI models in 2026