1stTop pick
Researched 11d agoQwen3.8-Max
- MMLU-Pro
- —
- Output (from)
- $6.00 / 1M
Best Small Language Models Under 10B Parameters (2026)
Last refreshed 2026-09-01. Next refresh: weekly.
The best small LLMs under 10B parameters in 2026 — fast, cheap, and deployable on-device or at the edge with strong benchmark scores.
Verdict
Use Qwen3.8-Max for small-model deployments today.
Jina Reranker v3.5 is the runner-up: — vs — on MMLU-Pro.
2ndShortlistRestricted
Researched 8d agoJina Reranker v3.5
- MMLU-Pro
- —
- Output (from)
- Free
3rdShortlist
Researched 59d agoLongCat-2.0
- MMLU-Pro
- —
- Output (from)
- $2.95 / 1M
Single-source resultMiniMax M2.7 scored 80.4% on MMLU-Pro, more than five points above the next GA score (56.0%). We dropped it one GA rank until another source corroborates the result.
How we rank
Small models (≤10B active parameters) rank on MMLU-Pro, then GPQA Diamond, MMLU, and HellaSwag.
- Eligibility — Non-deprecated models with ≤10B parameters (billions-only parser).
- Primary ranking — MMLU-Pro, then GPQA Diamond, then MMLU, then HellaSwag, then newer release.
- Podium freshness — Shortlist cards require `lastResearched` within 60 days and a tracked public output price. Stale or unpriced SKUs stay in the table with a “Verify pricing” badge once research is past 45 days.
- Variant collapse — We keep one row per model family (`familySlug` + parameter tier). When headline scores tie within ±0.5 pt (±10 Elo on Chatbot Arena), we pick the canonical SKU by lowest tracked input price, then GA over preview or limited access, then newest `release`. A folded sibling within the benchmark noise band can show a "Tied within margin" chip on that score cell.
- Pricing — SLMs often win on unit economics — compare the provider ladder before picking.
| # | Model | Input $/1M | Output $/1M | |
|---|---|---|---|---|
| 1 | Granite 4.1 8B MMLU-Pro: 55.99% | $0.05 | $0.10 | |
| 2 | MiniMax M2.7 ReasoningTools MMLU-Pro: 80.43% | $0.28 | $1.20 | |
| 3 | Phi-4 Mini MMLU-Pro: 52.8% | $0.90 | $0.90 | |
| 4 | Gemma 2 9B MMLU-Pro: 52.08% | $0.06 | $0.18 | |
| 5 | LFM2.5 8B A1B ReasoningTools MMLU-Pro: 50.5% | — | — | |
| 6 | Granite 4.1 3B MMLU-Pro: 49.83% | — | — | |
| 7 | Phi-3 Mini 4k MMLU-Pro: 45.66% | $0.05 | $0.25 | |
| 8 | LFM2.5 1.2B Instruct Tools MMLU-Pro: 44.35% | — | — | |
| 9 | Llama 3.1 8B Instruct MMLU-Pro: 44.25% | $0.02 | $0.05 | |
| 10 | Llama 3 8B Instruct MMLU-Pro: 40.5% | $0.02 | $0.04 | |
| 11 | Llama 3.2 3B Instruct MMLU-Pro: 34.7% | $0.03 | $0.05 | |
| 12 | Llama 3.2 1B Instruct MMLU-Pro: 20% | $0.03 | $0.10 | |
| 13 | Qwen3.8-Max ReasoningVisionTools MMLU-Pro: — | $2.00 | $6.00 | |
| 14 | Qwen3-8B MMLU-Pro: — | $0.04 | $0.14 | |
| 15 | Qwen2-7B MMLU-Pro: — | $0.05 | $0.15 | |
| 16 | Gemma 7B Instruct MMLU-Pro: — | $0.05 | $0.07 | |
| 17 | OpenChat 3.5 (0106) MMLU-Pro: — | $0.07 | $0.07 | |
| 18 | Starling LM 7B Beta MMLU-Pro: — | — | — | |
| 19 | Zephyr 7B Beta MMLU-Pro: — | $0.05 | $0.20 | |
| 20 | Qwen2.5-7B-Instruct MMLU-Pro: — | $0.03 | $0.03 |
Honorable mentions
Next seats in this ranking. Lines below are from each model's stored description in LLMReference seed data—spot-check the model page before relying on a capability claim.
Ornith-1.0 9B Dense is the smallest public variant of DeepReinforce's Ornith 1.0 family. Designed for edge deployment and single-GPU setups (~19 GB in BF16 on a single 80GB GPU). Post-trained on a Qwen 3.5 or Gemma 4 base using self-scaffolding RL for agentic coding tasks. The model learns to build and refine its own execution harness during RL rather than using a fixed scaffold. Supports OpenAI-compatible tool calls. Context: 262,144 tokens confirmed by the official Hugging Face model card and config.
—
MMLU-Pro
Higgs Audio v3 TTS is Boson AI's 4B-parameter text-to-speech model released June 4, 2026. It supports 102 languages (85 at production quality with WER/CER <5%), zero-shot voice cloning, and inline control tokens for emotion (21 types), style (singing/shouting/whispering), sound effects, and prosody. Audio output is 24kHz MP3 or PCM. Open weights available under a non-commercial license; hosted API is in free public preview.
—
MMLU-Pro
MOSS-TTS-v1.5 is an open-weight multilingual text-to-speech model from MOSI AI and the OpenMOSS team. The 8B-parameter MossTTSDelay model supports zero-shot voice cloning, long-form speech generation, explicit pause control with [pause X.Ys] markers, and language-tagged multilingual synthesis across 31 languages. Version 1.5 improves on MOSS-TTS v1.0 with stronger multilingual synthesis, more stable voice cloning, better long-reference short-text handling, and punctuation-driven prosody. The model weights are available on Hugging Face under Apache 2.0; no hosted token-priced API route is confirmed in the June 2026 research handoff.
—
MMLU-Pro
Compare Top Picks
Side-by-side comparison of the top picks by price, benchmark, and API access.
Browse Other Categories
Best LLMs for Code GenerationBest LLMs for RAGBest LLMs for Agents & Coding (2026)Best LLMs for ClassificationBest Open Source LLMsBest Multimodal / Vision LLMsBest LLM for Translation in 2026Best AI Image Models in 2026Best AI Video Models in 2026Best LLMs for Reasoning & MathBest LLMs for JSON / Tool useCheapest LLM APIs You Can Call Right NowBest Long Context LLMsBest Mainstream LLM APIs, RankedBest LLMs for EnterpriseBest Free LLMs You Can Use Right NowBest LLMs for WritingBest LLMs for MarketingBest LLMs for Customer Support