Gemma 4 26B A4B vs Step 3.7 Flash

Google · US  |  StepFun · China · Updated June 2026

Quick verdict

Pick Gemma 4 26B A4B for fast, cheap inference from a sparse moe (3.8b active of 25.2b total) or near-31b-dense quality at a fraction of the compute and memory-bandwidth cost. Pick Step 3.7 Flash for a 198b-parameter sparse moe vision-language model (11b active) built specifically for agentic coding and search workflows or scmp reported it 'outshines larger rivals' from deepseek and moonshot on some benchmarks despite its smaller active size. On a tight budget at scale, Gemma 4 26B A4B is the value pick.

Gemma 4 26B A4B (Google, US) and Step 3.7 Flash (StepFun, China) line up two different AI ecosystems against each other — a comparison that is as much about cost philosophy and openness as raw capability. Gemma 4 26B A4B is an Apache-2.0 open MoE with 25.2B total but only 3.8B active parameters, delivering near-31B-dense quality at a fraction of the inference cost. Step 3.7 Flash is stepFun's 198B MoE vision-language model built for coding and search agents - a distinct Chinese lab with genuinely competitive open weights. Their biggest split is price, and the breakdown below shows exactly how that plays out for your workload.

Key differences at a glance

Side-by-side specs

SpecGemma 4 26B A4BStep 3.7 Flash
ProviderGoogle (US) StepFun (China)
ReleasedApril 2, 2026 May 29, 2026
Context window256K (~393 pages) 256K (~393 pages)
Price (in/out)$0.12/$0.37 per 1M tokens $0.2/$1.15 per 1M tokens
Open weight?Yes — self-hostable Yes — self-hostable
Modalitiestext, image, video, code text, image, code
SWE-Bench VerifiedNot published Not published
MRCR v2 @ 1MNot published Not published

Who wins what

Fast, cheap inference from a sparse MoE (3.8B active of 25.2B total)

Gemma 4 26B A4B

At $0.12/$0.37 per 1M tokens it undercuts Step 3.7 Flash ($0.2/$1.15 per 1M tokens), and that gap compounds at volume.

Near-31B-dense quality at a fraction of the compute and memory-bandwidth cost

Gemma 4 26B A4B

An Apache-2.0 open MoE with 25.2B total but only 3.8B active parameters, delivering near-31B-dense quality at a fraction of the inference cost — and it runs cheaper at $0.12/$0.37 per 1M tokens.

Strong reasoning and coding (88.3% AIME 2026 no-tools, 77.1% LiveCodeBench v6)

Gemma 4 26B A4B

Gemma 4 26B A4B lists strong reasoning and coding (88.3% AIME 2026 no-tools, 77.1% LiveCodeBench v6) among its strengths; Step 3.7 Flash does not.

A 198B-parameter sparse MoE vision-language model (11B active) built specifically for agentic coding and search workflows

Step 3.7 Flash

Gemma 4 26B A4B is comparatively weak here — all 25.2B parameters must be loaded into memory even though only 3.8B are active per token

SCMP reported it 'outshines larger rivals' from DeepSeek and Moonshot on some benchmarks despite its smaller active size

Step 3.7 Flash

StepFun's 198B MoE vision-language model built for coding and search agents - a distinct Chinese lab with genuinely competitive open weights — and it is the newer of the two.

Open weights (Apache 2.0) at a low per-token price

Step 3.7 Flash

Step 3.7 Flash lists open weights (Apache 2.0) at a low per-token price among its strengths; Gemma 4 26B A4B does not.

Lowest cost at scale

Gemma 4 26B A4B

At $0.12/$0.37 per 1M tokens, it is the cheaper of the two — the gap dominates the bill on high-volume workloads.

Which should you pick?

A cost-sensitive startup shipping high volume

Gemma 4 26B A4B

At $0.12/$0.37 per 1M tokens it undercuts Step 3.7 Flash, and on millions of tokens that margin decides the monthly bill.

Anyone whose priority is fast, cheap inference from a sparse moe (3.8b active of 25.2b total)

Gemma 4 26B A4B

It is specifically built for that.

Anyone whose priority is a 198b-parameter sparse moe vision-language model (11b active) built specifically for agentic coding and search workflows

Step 3.7 Flash

That is its strongest area.

An enterprise with regional data-residency rules

Gemma 4 26B A4B or Step 3.7 Flash

Origin (US vs China) affects where data is processed and which compliance regime applies — check the provider's terms for your region.

Gemma 4 26B A4B: where it fits

An Apache-2.0 open MoE with 25.2B total but only 3.8B active parameters, delivering near-31B-dense quality at a fraction of the inference cost. Released April 2, 2026 by Google, it is built for fast, cheap inference from a sparse MoE (3.8B active of 25.2B total), near-31B-dense quality at a fraction of the compute and memory-bandwidth cost, strong reasoning and coding (88.3% AIME 2026 no-tools, 77.1% LiveCodeBench v6), and multimodal input (text/image, plus video processed as frames up to 60s) with native function calling.

Its trade-offs are real: all 25.2B parameters must be loaded into memory even though only 3.8B are active per token, and 256K context trails 1M-token frontier rivals, and this variant has no audio input (audio is E2B/E4B/12B only). At $0.12 in / $0.37 out per million tokens, it sits in the budget price band.

Step 3.7 Flash: where it fits

StepFun's 198B MoE vision-language model built for coding and search agents - a distinct Chinese lab with genuinely competitive open weights. Released May 29, 2026 by StepFun, it is built for a 198B-parameter sparse MoE vision-language model (11B active) built specifically for agentic coding and search workflows, sCMP reported it 'outshines larger rivals' from DeepSeek and Moonshot on some benchmarks despite its smaller active size, open weights (Apache 2.0) at a low per-token price, and includes a 1.8B vision encoder for image understanding alongside text.

Its trade-offs: stepFun is a newer, less established lab than DeepSeek, Alibaba, or Moonshot, benchmark 'outshines rivals' claims are from press coverage of specific tests, not a full independent leaderboard sweep, and smaller ecosystem and less third-party documentation than the more established Chinese labs. At $0.2 in / $1.15 out per million tokens, it sits in the budget price band.

The bottom line for this matchup

This is less "which is smarter" and more "which ecosystem fits." Gemma 4 26B A4B (US) and Step 3.7 Flash (China) differ on pricing philosophy, data-residency, and tooling as much as on raw scores. Gemma 4 26B A4B is the cheaper option, which matters at volume. The pragmatic move is to run one real task through both and judge the outputs against your own constraints — including where your data is allowed to be processed.

Want both Gemma 4 26B A4B and Step 3.7 Flash without two subscriptions? LumiChats gives you these plus 40+ models under one ₹69/day pass (about $1/day) — draft with one, cross-check with the other.

See pricing

Frequently asked questions

Is Gemma 4 26B A4B or Step 3.7 Flash better for coding?

Public SWE-Bench figures are not available for either model, so the honest test is your own repository — run an identical real bug through both. By design, Gemma 4 26B A4B leans toward fast, cheap inference from a sparse moe (3.8b active of 25.2b total) while Step 3.7 Flash leans toward a 198b-parameter sparse moe vision-language model (11b active) built specifically for agentic coding and search workflows, and that positioning usually predicts which feels better on your codebase.

Which is cheaper, Gemma 4 26B A4B or Step 3.7 Flash?

Gemma 4 26B A4B is cheaper — $0.12/$0.37 per 1M tokens vs $0.2/$1.15 per 1M tokens, roughly 1.7× apart on input.

Which has the bigger context window?

Both advertise 256K (~393 pages). Remember advertised ≠ usable: recall typically degrades before the ceiling.

Can I use both Gemma 4 26B A4B and Step 3.7 Flash together?

Yes — a multi-model platform like LumiChats gives you Gemma 4 26B A4B, Step 3.7 Flash and 40+ others under one ₹69/day pass (about $1/day), so you can draft with one and cross-check with the other instead of buying two subscriptions.

Which is newer, Gemma 4 26B A4B or Step 3.7 Flash?

Step 3.7 Flash — released May 29, 2026, about 57 days after Gemma 4 26B A4B.

Related comparisons

Specifications and benchmarks reflect publicly reported figures as of June 2026 and may change as providers release updates. Always verify on your own workload.