explainx / blog
Cerebras runs Gemma 4 31B at 1,851 tokens/sec — 35× GPU speed, first multimodal on wafer-scale hardware. Haiku-class intelligence, Apache 2.0, screenshot agents.
Jun 4, 2026
Google DeepMind's Gemma 4 12B is a breakthrough open-source model that delivers flagship-level multimodal capabilities on consumer hardware. Here's everything you need to know about architecture, benchmarks, deployment, and use cases.
Jun 29, 2026
Zhipu founder Jie Tang's June 29 poll drew 466K views. Vision, shorter thinking, and llama.cpp day-one support top the list — GLM-5.2 is text-only; users want Opus-class multimodal next.
Jun 30, 2026
InternScience's Agents-A1 claims trillion-class agent performance at 35B MoE scale — BrowseComp 75.5, GAIA 96.0, IFEval 94.8 — with open weights on Hugging Face and ModelScope. Here is what the paper claims, how to serve it, and the honest gaps on coding and Qwen lineage.
June 29, 2026: Cerebras announced Gemma 4 31B running at 1,851 output tokens per second on Cerebras Inference — 35× a typical GPU endpoint per Artificial Analysis benchmarking. It is the first Google DeepMind model on the platform and the first multimodal model at wafer-scale speed: developers can feed images — screenshots, documents, charts, UI states — into inference that previously only text models achieved at this throughput.
For teams building visual agent loops while Fable 5 remains offline and GLM-5.3 vision is still a community wishlist, Gemma 4 on Cerebras is a different bet: open-weight multimodal at real-time latency, not frontier closed API access.
| Item | Detail |
|---|---|
| Model | Gemma 4 31B — Google DeepMind flagship dense open model |
| Speed | 1,851 output TPS (Artificial Analysis); 35× typical GPU |
| First token | 1.5 s inclusive of reasoning |
| Intelligence | AA Index 29 — comparable to Claude Haiku 4.5 (30) |
| vs Haiku speed | 18× faster on Cerebras |
| License | Apache 2.0 — open weights |
| Multimodal | First on Cerebras — images + text |
| Availability | Cerebras Inference Cloud — public preview (limited time) |
| Local alt | Gemma 4 12B on 16GB VRAM |
Cerebras has benchmarked fast inference across open-weight stacks — Kimi, GLM, GPT-OSS, Qwen — on wafer-scale hardware. Gemma 4 is the first Google DeepMind model on the platform, and the first where vision enters the loop at Cerebras speed.
Olivier Lacombe, Product Lead for Gemma at Google DeepMind:
"Gemma 4, Google DeepMind's family of open models, was built to bring advanced reasoning and multimodal capabilities at developer-friendly sizes. Pairing these capabilities with Cerebras's wafer-scale technology provides developers with an exciting platform for running extremely fast visual and agentic workflows."
The pitch is not "same product, faster." Logan Kilpatrick (Google DeepMind):
"If every model was doing 2,000 tokens per second, you would probably build different products. You wouldn't build the same product and just have it be faster."
At 1,800+ TPS, multimodal agent loops — inspect image → reason → structured output → tool call → verify → retry — stop feeling like batch jobs and start feeling interactive.
| Metric | Gemma 4 31B on Cerebras | Typical GPU endpoint |
|---|---|---|
| Output TPS | 1,851 | ~53 (35× slower) |
| First token (incl. reasoning) | 1.5 s | Much higher on GPU |
| vs Claude Haiku 4.5 (same hardware class) | 18× faster | — |
| Intelligence (AA Index) | 29 | Haiku 4.5: 30 |
Cerebras recommends Gemma 4 31B as the reference medium-size model on its cloud: an alternative to Haiku, GPT-OSS, or Llama with equal or higher intelligence at Cerebras speed.
Dense vs MoE: Gemma 4 31B is a dense multimodal model — high intelligence without the large memory footprint of MoE serving. That fits Cerebras's wafer-scale serving story: strong enough for serious agent work, efficient to run at scale, Apache 2.0 for build-around freedom.
Before Gemma 4, Cerebras Inference was text-first at extreme speed. Gemma 4 adds:
Cerebras states multimodal support starts with Gemma 4 and will extend to additional models — the same platform pattern they used for Kimi and GLM text inference.
This directly addresses the workflow gap Zhipu users asked for in GLM-5.3 polls: vision integrated with reasoning, not a separate VL model bridge. Gemma 4 on Cerebras ships that integration today — on cloud wafer hardware, not local 16GB VRAM.
For on-device multimodal, see Gemma 4 12B local guide and Gemma 4 E4B + Argent iOS automation.
Feed a dense dashboard screenshot or document page. The model identifies what matters, explains the finding, returns structured output — in real time, not after a GPU wait.
Hand it a research report or technical brief. Get a decision-ready summary fast enough to read, react, and re-query in one sitting — relevant for teams comparing against Kimi K2.7's long-context coding or GLM document pipelines.
Play to medium-model strengths: broken UI screenshot + source + console error → minimal patch + verification checks. This is the agentic coding loop Fable 5 marketed — now runnable on open weights at speeds that keep a human in the loop.
Gemma 4's multimodal stack supports UI state reasoning — overlapping with E4B + Argent simulator navigation at the edge, but at cloud scale and speed for heavier agents.
Multimodal and agentic workflows rarely call a model once:
Visual input → reason → structured output → tool call → check result → retry
At conventional GPU speeds, each loop adds seconds — too slow for real-time copilots, live UI repair, or tight human-in-the-loop review.
At 1,800+ TPS:
The product implication is architectural: you design for continuous visual reasoning, not "send screenshot, wait, hope."
| Dimension | Gemma 4 31B (Cerebras) | Claude Haiku 4.5 |
|---|---|---|
| Intelligence (AA Index) | 29 | 30 |
| Speed on Cerebras | 1,851 TPS | 18× slower (per Cerebras) |
| License | Apache 2.0 open weights | Proprietary API |
| Export control | No US nationality gate | Anthropic API terms |
| Multimodal | Native on Cerebras | Native via Anthropic API |
| Fable-class coding | Medium model — not Fable tier | Haiku ≠ Fable; Fable still offline |
For international developers blocked from Fable 5, Gemma 4 + Cerebras is another unrestricted multimodal path — alongside GLM-5.2 text and Kimi K2.7-Code coding.
Haiku remains marginally smarter on the AA Index. Gemma 4 wins on speed, openness, and self-host/build-around freedom when paired with Cerebras cloud.
Cerebras has positioned itself as the inference speed leader for open weights:
If your stack already routes GLM-5.2 through Z.ai for text coding, Gemma 4 on Cerebras is the vision complement — similar to the Qwen-VL → GLM bridge developers want eliminated in GLM-5.3, but with single-model multimodal at extreme speed.
Gemma 4 31B is on the Cerebras Inference Cloud in public preview for a limited time as of June 29, 2026.
Cerebras asks teams with workloads in:
…to contact them directly via cerebras.ai.
This is not a self-serve unlimited free tier announcement — preview access with enterprise outreach for heavy multimodal pipelines.
Evaluate Cerebras Inference Cloud preview for Gemma 4 31B. Benchmark your screenshot-to-patch or dashboard-insight loop against GPU baselines — Cerebras claims 35× on output TPS.
Use Gemma 4 12B locally — 16GB VRAM, Apache 2.0, unified architecture. Trade wafer-scale TPS for data never leaving your machine.
GLM-5.2 and Kimi K2.7-Code remain the open-weight coding leaders while Fable is suspended.
Compare Gemma 4 on Cerebras (cloud speed) vs Gemma 4 E4B + Argent (local iOS simulator) — same family, different deployment surface.
Is Gemma 4 on Cerebras the fastest multimodal inference available?
Per Cerebras and Artificial Analysis methodology cited in the announcement — yes, as of June 29, 2026, at 1,851 output TPS for Gemma 4 31B.
Does it replace Fable 5 or Opus for coding?
No — it targets Haiku-class medium intelligence with multimodal speed. Serious autonomous coding at Fable depth still points to restricted US models or open alternatives (Kimi, GLM, LongCat-2.0).
Does it matter for product design?
Yes. Kilpatrick's quote is the thesis: at 2,000 TPS, you build different products — not the same agent with shorter waits. Multimodal loops become real-time collaborators.
Speed and intelligence figures cite Cerebras's June 29, 2026 announcement and Artificial Analysis benchmarking as referenced by Cerebras. Preview availability and pricing may change — verify on cerebras.ai before production commitments.