Gemma 4 (Google)
A Tier · 8.3/10
Google DeepMind's open-weights model family -- multimodal, 256K context, runs on edge devices **EmbeddingGemma 2 (announced 2026-10-06; weights on Hugging Face since 9/14) adds a 740M-parameter Apache 2.0 embedding model built on Gemma 4 that maps text, code, images, video and audio into one 768-dimensional space, with an 8K context, Matryoshka truncation down to 128 dimensions and about 191 MB of RAM for text-only use on a Pixel 11 Pro.**
Score Breakdown
Benchmark Scores
Benchmarks for Gemma 4 31B
| Benchmark | Description | Score | |
|---|---|---|---|
| MMLU | Knowledge across 57 subjects | 83% | |
| GPQA Diamond | Graduate-level science questions | 84.3% | |
| AIME 2026 | 89.2% | ||
| HumanEval | Python code generation | 85% |
Last updated: 2026-04-13
Personality & Tone
The compact Google cousin
Tone: Similar corporate-Google tone as Gemini but smaller and less polished. Gemma's chat replies are short, cautious, and structured -- closer to a careful intern than a peer.
Quirks: Inherits a Gemini-like safety bias, so refusals appear on prompts Mistral or DeepSeek would answer. Best used as a cheap local fallback or on-device model, not as a personality play.
The Good and the Bad
What we like
- +Apache 2.0 license -- truly permissive, you can use it commercially without strings attached
- +Multimodal: handles text + image input (audio on smaller models), generates text output
- +256K token context window -- larger than most open models
- +140+ language support -- one of the strongest multilingual open models available
- +Four sizes (E2B, E4B, 26B MoE, 31B Dense) cover edge devices to data centers
- +31B Dense scores 89% on AIME 2026 and 84% on GPQA Diamond -- competitive with frontier closed models
- +31B Dense currently ranks #3 among open models on the LMArena text leaderboard (26B MoE ranks #6) -- genuinely competitive with the top Chinese and Meta open-weight families
- +26B MoE activates only 3.8B params during inference for fast tokens-per-second
What could be better
- −Requires technical setup unless you use a hosted API provider
- −Quality still trails the very best closed models (GPT-6 Astra, Claude Opus 5.5, Gemini 3.8) on hardest reasoning tasks
- −No native chat UI from Google -- you're either coding against an API or using a third-party frontend
- −Smaller community than Llama -- fewer fine-tunes and tooling integrations exist
Pricing
Self-hosted
- ✓Apache 2.0 license
- ✓Free download from Hugging Face/Kaggle/Ollama
- ✓Run on your own hardware
API (OpenRouter, Gemma 4 31B)
- ✓Hosted inference
- ✓$0.14 input / $0.40 output
- ✓No infrastructure setup
Google AI Studio
- ✓Free tier for testing
- ✓Web playground access
System Requirements
Hardware needed to self-host. Min = smallest viable setup (usually heavy quantization). Max = full-precision / production-grade.
| Model variant | Min | Max |
|---|---|---|
| Gemma 4 E2B / E4B (edge-class) | 2-3 GB VRAM Q4 (runs on phones and laptops) | 8-12 GB VRAM FP16 |
| Gemma 4 26B MoE | 8 GB VRAM (Q4) | 32 GB VRAM FP16 |
| Gemma 4 31B Dense (flagship) | 12 GB VRAM Q4 (RTX 4070) | 1× A100 40 GB FP16 |
Known Issues
- EMBEDDINGGEMMA 2 -- A 740M OPEN MULTIMODAL EMBEDDING MODEL ON THE GEMMA 4 ARCHITECTURE (2026-10-06 blog; Hugging Face repo created 2026-09-14; vendor-primary): Google announced **EmbeddingGemma 2**, 'the most capable model for on-device multimodal embeddings', which maps 'text (incl. code), images, video, and audio inputs -- and combinations thereof -- into a single, unified 768-dimensional vector space'. **Specs (model card):** 740M total parameters = a 270M text model (130M transformer + 140M embedder) plus selectively loadable vision (170M) and audio (300M) encoders; **Apache 2.0** (the card states Apache 2.0 while linking Google's Gemma 4 license page); 100+ languages; **8K token context** ('4x larger than EmbeddingGemma 1', enough for about 5.5 minutes of audio, 29 images or 58 video frames); **Matryoshka Representation Learning** lets you truncate 768-d vectors to 512, 256 or 128 for up to 6x storage reduction; task-steered instruction prefixes. **On-device footprint:** with quantization on a Pixel 11 Pro, about 191 MB of active RAM for text-only weights and about 567 MB for the full multimodal model. **Vendor benchmarks:** MTEB Code 78.68 (from 68.76 for EmbeddingGemma 1, a 9.92-point gain), 'leading scores among sub-1B multimodal embedders' on MTEB Code and MAEB, 'matching or outperforming many larger models'. Because it 'shares its text tokenizer and audio encoder' with Gemma 4, Google pitches a combined on-device RAG stack with a lower total memory footprint. Context: the first EmbeddingGemma passed 20 million downloads. **Timing note:** the Hugging Face repo google/embeddinggemma-2 was created 2026-09-14, three weeks before the blog post -- weights before write-up, the pattern this site has seen with Qwen. Not a Gemma 4 LLM revision; recorded here because this is the Gemma family page and the model is the retrieval half of Google's on-device open stack.Source: Google (blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/, root RSS pubDate Tue, 06 Oct 2026, on-page 'Oct 06, 2026'); Hugging Face model card (huggingface.co/google/embeddinggemma-2/raw/main/README.md -- 'license: apache-2.0', 740M, 768-d, 8K; API createdAt 2026-09-14) -- both fetched 2026-10-06 via curl · 2026-10-06
- 2026-04-18 BF16 stability refresh -- Google re-released Gemma 4 multimodal checkpoints in BF16 format focused on truthfulness, JSON / tool-call formatting, long-context extraction reliability, and loop resistance. Not a new model version; a quality refresh that fixes specific failure modes developers were hitting in production. If you pulled weights before 2026-04-18, consider re-downloading for the new checkpointsSource: Google DeepMind Gemma page, Hugging Face · 2026-04
- Gemma 4 launched April 2, 2026 with improved licensing -- earlier Gemma versions had restrictive use clauses that confused developersSource: The Register, Hugging Face · 2026-04
- Function calling support is new -- some users report inconsistent tool-use behavior compared to Llama 3 or MistralSource: Hugging Face discussions · 2026-04
Best for
Developers and businesses who need a permissively licensed multimodal LLM they can self-host or fine-tune. Especially good for multilingual use cases and on-device deployment.
Not for
Non-technical users who just want to chat with an AI -- there's no consumer-facing app. Use Gemini if you want a polished chat experience.
Our Verdict
Gemma 4 is Google's answer to the open-weights race against Meta's Llama and the wave of strong Chinese open models. The Apache 2.0 license is a big deal -- it removes the legal friction that made earlier Gemma adoption awkward. The 31B Dense model is genuinely competitive with frontier closed models on benchmarks while costing $0.14/M input via API. If you're building a product on open-weights LLMs and you need multimodal + multilingual + permissive licensing, Gemma 4 is now a top choice.
Sources
- Google: EmbeddingGemma 2 -- an open, lightweight multimodal embedding model (2026-10-06) (accessed 2026-10-06)
- Hugging Face: google/embeddinggemma-2 model card -- Apache 2.0, 740M, 768-d, 8K context (repo created 2026-09-14) (accessed 2026-10-06)
- Google DeepMind Gemma 4 page (accessed 2026-04-08)
- Google blog: Gemma 4 launch (accessed 2026-04-08)
- Artificial Analysis benchmarks (accessed 2026-04-08)
- OpenRouter pricing (accessed 2026-04-08)
- The Register coverage (accessed 2026-04-08)
Explore more Gemma 4 (Google) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for Gemma 4 (Google).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to Gemma 4 (Google)
Llama 4 (Meta)
Meta's open-weights family -- Scout (10M context), Maverick (multimodal 400B MoE). NOTE: Meta's frontier work moved to the proprietary Muse Spark line in April 2026; Llama remains downloadable and supported but is effectively in maintenance mode
Mistral AI
**Mistral Large 4 ('le Chonk', ML4) entered public preview on 2026-10-06: a 1T-parameter natively multimodal MoE with roughly 50B active parameters, a 1M context window, callable today as mistral-large-4 at $1.36 / $0.14 cached / $4.18 per 1M tokens (batch at half price), open weights promised by the end of October, and vendor-claimed open-weight leads in cyber (Cybench 93%), coding (DeepSWE v1.1 61.7%) and agentic work (AutomationBench 59.9%).** **Mistral opened a Munich hub for Physics AI and Industrial AI on 2026-09-28**, confirming the May 2026 acquisition of Emmi AI (30+ physicists and engineers in computational fluid dynamics, structural mechanics and multi-physics simulation), naming BMW (crash simulation) and Siemens Energy as partners, and committing to **build one gigawatt of European compute capacity by 2030**. **Mistral now powers Firefox Smart Window (2026-09-16)** -- Mozilla's beta AI browsing assistant runs on Mistral models for France and North America, UK and Germany later this year, with zero data retention. **EUR 3B Series D at a EUR 21B+ post-money (9/08, Samsung-led)**; Cloudera partnership (9/10); Medium 3 retired 8/31 in favour of Medium 3.5.
DeepSeek
**DeepSeek-V4.1-Flash shipped 2026-09-10**: a 552B MoE on a new Causal Encoder-Decoder architecture (8B active for input, 16B for output), natively multimodal, MIT-licensed on HF, and priced BELOW the V4-Flash it replaces -- $0.15/$0.60 off-peak, $0.30/$1.20 peak per 1M, with cache-hit input cut to $0.003/$0.006. V4-Flash and V4-Flash-Vision-Exp are retired (aliases route to V4.1-Flash). DeepSeek said V4-Pro would be phased out on 9/14, then **reversed that four days later** -- V4-Pro stays on the API at unchanged $0.66/$1.98 off-peak, $1.32/$3.96 peak until V4.1-Pro lands. V4.1-Flash's vendor benchmarks beat V4-Pro (DeepSWE 74.2, Terminal-Bench 2.1 90.6, GPQA 90.9)
DiffusionGemma (Google)
Google DeepMind's experimental open-weights TEXT-DIFFUSION model (June 10, 2026) -- 26B MoE (3.8B active), Apache 2.0, generates 256-token blocks in parallel with bidirectional attention for up to 4x faster output (1,000+ tok/s on H100). Trades some quality vs Gemma 4 for raw speed
Qwen (Alibaba)
**Qwen-Image-2.1 went open-weight (weights 2026-09-14, blog 2026-09-20)** -- a unified text-to-image and editing model with a 7B visual-generation component, native transparent (RGBA) output, up to 10 reference images and mask-guided local edits -- but under a **research-only licence, not Apache**: commercial use needs a separate licence from Alibaba. Alibaba's open-weights + API family, and August 2026 was its biggest open-release month yet: **Qwen3.8-2.4T-A95B -- the Max-class flagship -- went open-weight on 2026-08-08** (custom Qwen3.8-Max licence), **Qwen3.8-27B** dense landed Apache 2.0 (8/05), and **Qwen3.8-Flash-Next** (8/24 weights, 8/26 blog) previews the Qwen4 architecture: 125B main + 51B n-gram embeddings with only 6B active, served as Qwen3.8-Flash at $0.15/$0.47 per 1M. Qwen 3.7 Max remains the GA API flagship ($2.50/$7.50)
GLM / Z.ai (Zhipu AI)
**GLM 5.2 has a named production deployment at Perplexity: on 2026-09-21 Perplexity disclosed that Perplexity Computer sessions are served by GLM 5.2, post-trained by Perplexity on real user sessions** (21.2% relative cut in live tool-call failures). **GLM-5.3 and GLM-5.3-Flash released 2026-08-25** (weights on Hugging Face): 5.3 is a post-training-only upgrade of the 5.2 base that Z.ai calls 'the most capable open-weights model for coding' (Terminal Bench 2.1 88.2, DeepSWE 66.9, CyberGym 84.5) but moves from MIT to a custom licence; 5.3-Flash (320B/18B active, multimodal) stays MIT at $0.15/$0.50 per 1M.
Kimi K3 (Moonshot)
Moonshot's 2.8T-parameter Kimi K3 (launched 2026-07-16/17) is the largest open-weight model ever released -- 1M context, multimodal, $3/$15 per 1M via API, ranked best-available on Arena.AI at launch. WEIGHTS SHIPPED ~2026-07-26/27 on Hugging Face (2.8T total / 104B activated, safetensors) under a custom Kimi K3 License, not the Modified MIT of the K2 line
Nemotron (Nvidia)
Nvidia's open-weights family -- hybrid Mamba-Transformer MoE architecture, optimized for efficient reasoning on Nvidia hardware. Nemotron 3 Ultra (550B total / 55B active) shipped 2026-06-04 as the family flagship, joining Super (120B/12B, March) and Nano
MiniMax M3
**MiniMax-H3's weights are public on Hugging Face (repo 2026-07-28, updated 2026-08-13) under a community licence whose open-weight grant is limited to the US, EU, UK and South Korea**, and **MiniMax Music 3 (2026-08-07)** ships open weights for five-minute full-song generation with a prominent-attribution commercial clause. MiniMax's coding/agent flagship -- M3 (June 1 2026): 1M-token context, MSA sparse attention (>15x decoding speedup at long context), SWE-Bench Pro 59.0%, Terminal-Bench 66.0%. OPEN WEIGHTS LIVE on HuggingFace since June 12 (~428B total / ~23B active, native multimodal, minimax-community license)
Falcon (TII)
UAE's Technology Innovation Institute open-weights family -- Falcon 3 optimized for efficient sub-10B deployment on consumer hardware
gpt-oss (OpenAI)
OpenAI's FIRST open-weight models -- gpt-oss-120b (single 80GB GPU, near parity with o4-mini on reasoning) and gpt-oss-20b (runs on 16GB edge devices). Apache 2.0. Launched 2025-08-05. gpt-oss-safeguard ships in 2026 as the safety-tuned variant
IBM Granite 4.0
IBM's enterprise-focused open-weight family -- Granite 4.0 hybrid Mamba-2 + transformer architecture (70-80% memory reduction vs pure transformer), 3B to 32B sizes, Apache 2.0. First open model family to secure ISO 42001 certification. Nano 350M runs on CPU with 8-16GB RAM. 3B Vision variant landed 2026-04-01
Arcee Trinity-Large-Thinking
Arcee AI's US-made open-weight frontier reasoning model -- launched 2026-04-01. 398B total params, ~13B active. Sparse MoE (256 experts, 4 active = 1.56% routing). Apache 2.0, trained from scratch. #2 on PinchBench trailing only Claude 3.5 Opus. ~96% cheaper than Opus-4.6 on agentic tasks
Olmo 3 (AI2)
Allen Institute for AI's fully-open frontier reasoning models -- Olmo 3 family (2025-11-20) includes 7B and 32B sizes, four variants (Base, Think, Instruct, RLZero). Apache 2.0 with fully open data + checkpoints + training logs. Olmo 3-Think 32B matches Qwen3-32B-Thinking at 6x fewer training tokens
AI21 Jamba2
AI21 Labs' hybrid SSM-Transformer (Mamba-style) open-weight family -- Jamba2 launched 2026-01-08. Two sizes: 3B dense (runs on phones / laptops) and Jamba2 Mini MoE (12B active / 52B total). Apache 2.0, 256K context, mid-trained on 500B tokens
StepFun Step 3.7 Flash
StepFun's (China) agent-focused open-weight family -- Step 3.7 Flash (May 28 2026): 198B sparse MoE vision-language model, ~11B active, 256K context, Apache 2.0, ~400 tok/s, SWE-Bench Pro 56.3. Supersedes Step 3.5 Flash (Feb 2026) as the flagship
Cohere Command A
Cohere's enterprise-multilingual flagship -- 111B params, 256K context, runs on 2x H100. 23 languages. CC-BY-NC 4.0 on weights (research / non-commercial), commercial requires Cohere enterprise contract. Follow-ups: Command A Reasoning + Command A Vision
LongCat-2.0 (Meituan)
Meituan's open-source 1.6T-parameter MoE (~48B active) with native 1M-token context, MIT license -- trained entirely on domestic Chinese AI ASICs and revealed as the stealth 'Owl Alpha' model that had been topping OpenRouter
Inkling (Thinking Machines Lab)
Mira Murati's $12B lab ships its first model (2026-07-15): a 975B/41B-active open-weights MoE that reasons natively over text, images, and audio with a 1M-token context -- positioned not as the strongest model, but as the best starting point for fine-tuning via Tinker
Bonsai 27B (PrismML)
The first 27B-class model that runs on a phone (2026-07-14) -- ternary and 1-bit quantizations of Qwen3.6 27B squeeze a multimodal, tool-calling, 262K-context model into 3.9-5.9GB under Apache 2.0