Cerebras added Qwen 3.8 27B to its public inference tier running at roughly 1,500 tokens per second — a speed no GPU cluster matches at this price point, made possible by Cerebras's wafer-scale hardware keeping the entire model resident in on-chip SRAM rather than shuttling weights across memory bandwidth. The launch replaced Gemma 4 31B on the same shared, pay-as-you-go tier the same day, with no announced migration window.
The raw number is real, and it's the reason this launch shot to the top of Hacker News within hours. But the discussion underneath the headline is where the more useful story is: a token-per-minute cap that counts cached tokens at full price, and a context ceiling that together change who this is actually a good deal for. For background on how token throughput and context windows interact in agentic workloads, see explainx.ai's LLM context window explained and model selection and energy math.
TL;DR
| Question | Short answer |
|---|---|
| How fast is it? | Roughly 1,500 tokens/second on Cerebras's public endpoint |
| What replaced what? | Qwen 3.8 27B replaced Gemma 4 31B on Cerebras's shared, pay-as-you-go tier, same day |
| What's the rate limit? | 150,000 tokens per minute — and that includes cached/repeated input tokens |
| Is caching discounted? | No — Cerebras bills cached and fresh input tokens at the same standard rate |
| What's the context limit? | 128k tokens on this endpoint |
| Best fit? | Short, bursty, low-concurrency prompts — not sustained agentic coding sessions |
| Worst fit? | Long-running agentic workflows that resend large context on every tool call |
The speed is genuinely unusual
Cerebras's Wafer-Scale Engine keeps model weights entirely in on-chip SRAM instead of external HBM, which is the structural reason it can push raw generation speed an order of magnitude past typical GPU-hosted inference for a model this size. Multiple early testers on Hacker News independently clocked it around 1,500 tokens/second in practice, consistent with Cerebras's own published figures. One tester summarized the split cleanly: "Output is awesome, super fast as you expect from the 1500t/sec... Input doesn't look faster than other models." That distinction matters — the speed advantage is concentrated in generation, not in ingesting the prompt itself.
For workloads that are latency-sensitive and bursty — a single classification call, a quick tool-use round-trip, an interactive feature where every hundred milliseconds is visible to a user — that's a meaningful, differentiated advantage no other hosted provider currently matches at comparable pricing.
The rate limit and the caching gap
The complications start with Cerebras's public-tier rate limit: 150,000 tokens per minute. On its own, that sounds generous. In practice, several developers hit it within 90 seconds to a few minutes of real agentic use, because the limit counts all input tokens, including tokens served from cache, at the standard rate.
That second detail is the one that changes the economics. Cerebras's own documentation states plainly: "Input tokens, whether served from the cache or processed fresh, are billed at the standard input token rate for the respective model." Most inference providers — OpenRouter aggregating multiple backends among them — discount cache hits by 75-90%, because agentic coding workflows characteristically resend a large, mostly-unchanged context on every single turn (system prompt, prior conversation, file contents, tool definitions). Without a cache discount, that repeated context gets billed at full price every time, and burns against the 150k TPM cap every time too.
One tester's math made the effect concrete: a short coding review session with a handful of follow-ups, at a 91.4% cache hit rate on the underlying conversation, still cost $1.60 and took just over five minutes on Cerebras — versus an estimated $0.29 and about 14 minutes using typical OpenRouter provider averages for a comparable model. Cerebras was roughly 2.8x faster in wall-clock time, but 5.6x more expensive for the same session. Whether that trade is worth it depends entirely on whether your bottleneck is developer wait time or token spend — for high-volume, sustained agentic work, most users concluded it isn't.
Context window: 128k, not the model's native ceiling
Qwen 3.8 27B's own weights natively support a considerably longer context window, extensible even further with standard scaling techniques. Cerebras's hosted endpoint caps it at 128k tokens. For a single-turn or short-session use case, that's ample. For sustained agentic coding — the workload most people actually want blazing generation speed for — 128k gets consumed quickly once you account for system prompt, tool definitions, file contents, and the running conversation history, and several testers reported the model losing earlier context well before reaching the cap in real multi-file coding tasks.
Is Qwen 3.8 27B itself any good?
Separate from the hosting economics, the model drew consistently positive reports on capability. One developer running it locally, quantized, on consumer hardware put it plainly: "27B is rather smart. As many (or more) activated parameters as the flash models of the universe that we know about. It reasons very well. It just doesn't have a lot of knowledge." That "narrow knowledge, strong reasoning" profile is a common pattern for smaller, reasoning-tuned models, and it maps onto why several testers paired it with local search or retrieval rather than relying on it as a standalone knowledge source.
Reports were not uniformly positive — at least one tester running it self-hosted described it looping and losing track of its own task, while another running what appears to be a similar configuration called the same base model "unbelievably cracked and dedicated," successfully one-shotting a toy compiler with a JIT backend. The gap between those two experiences points to deployment configuration (quantization level, sampling parameters, reasoning-effort setting) mattering as much as the base model weights for this specific model.
What people are asking
Should I switch my agentic coding setup to Cerebras for this? Only if your sessions are short and your value-per-minute of saved wait time is high enough to offset a roughly 5-6x cost multiplier versus typical providers. For anything resembling a long agentic session with repeated tool calls, the missing cache discount and the 150k TPM ceiling work directly against you.
Why doesn't Cerebras discount cached tokens? The company hasn't given a definitive technical explanation, but several commenters speculate it relates to the architecture: Cerebras's wafer-scale design may not persist KV cache state between requests the way GPU-based inference stacks typically do, meaning "cached" here may functionally mean "we recognize this content," not "we skip recomputing it." That would explain both the lack of a discount and the absence of any latency benefit from repeated context.
Is this a good deal for anything other than coding? Yes — one-shot classification, quick lookups, tool-call orchestration with short context, and latency-critical interactive features are all workloads where Cerebras's speed advantage translates directly into user-facing value without the token-per-minute ceiling becoming the binding constraint.
Does the shared tier being removed without notice matter for production use? It should factor into anyone's decision to build on it. Cerebras's own team has previously characterized the shared, pay-as-you-go tier as not intended for production workloads — Gemma 4 31B's same-day removal with no migration window is consistent with that framing, and anyone building against Cerebras's shared tier should plan for models being swapped or removed without advance notice.
Related reading on explainx.ai
- LLM Context Window Explained (2026) — background on why context limits matter more than raw token counts for agentic work
- Model Selection and AI Energy Math (2026) — how to weigh speed, cost, and capability tradeoffs across providers
- Qwen3.8 Flash and Qwen4 Architecture Preview — where this model sits in Alibaba's broader Qwen 3.8 lineup
- GLM 5.3 Flash: Pricing, Fable 5.1, SimpleBench — another recent fast/cheap model release for comparison
- Nemotron 3.5 Lightning 30B-A3B: Open MoE — a similarly speed-focused open-weight release from NVIDIA
- Recursive Reasoning and Inference-Time Scaling — background on how reasoning-tuned smaller models trade knowledge for reasoning depth
- What Is Multimodal AI? Complete Guide — background context on the broader open-weight model landscape
Official source: Cerebras Qwen 3.8 27B model page
Pricing, rate limits, and throughput figures in this post reflect Cerebras's public documentation and early user reports as of September 4, 2026 and may change without notice, consistent with Cerebras's prior same-day model swaps on this tier.