Open-Source LMArena Rankings: 7 Models Closing the Gap

OLMo and GLM closed the gap on Claude. Then enterprise procurement teams read the licensing fine print — and four of the top-10 open-weight models become unusable overnight. That's not a marketing line. That's the average outcome of a real Q1 2026 vendor-risk review on open-source LMArena leaders.

This page is the procurement-grade decode of the open-weight tier — which 7 models genuinely close the proprietary gap, where they sit on the leaderboard, and which licensing clauses turn the headline rank into a deployment blocker. For the broader cross-cluster context and the live top-10 widget, start at the parent pillar: LMSYS/LMArena Top Models 2026 pillar. This sub-page zooms specifically into what enterprise architects actually need before signing off on an open-weight deployment.

The official source we cross-reference throughout is the live LMArena leaderboard at lmarena.ai — verify any procurement-grade decision against it directly before any infrastructure commitment.

The Open-Weight Top-7 — Where They Actually Rank

Seven open-weight models now hold credible LMArena positions in May 2026. Approximate Text leaderboard Elo (rounded, ±95% CI):

That's a 42-Elo-point gap from #1 (Claude Opus 4.6 at 1504) down to #10 GLM-4.7 — meaningful, but not disqualifying for most workloads. Two years ago that gap was 95+ points.

For the full month-over-month picture and the specific January 2026 vote-pipeline overhaul that reshuffled these rankings, see our companion: April 2026 LMArena Shake-Up: 3 Models Crashed Out of the Top-10.

Why GLM-4.7 Became the New Open-Weight Reference

GLM-4.7's late-March 2026 entry into the top-10 on both Text and WebDev simultaneously was the first time any open-weight model achieved that. The reasons are structural, not coincidental.

The capability story:

• Text Elo 1462 (within 42 points of Claude Opus 4.6, the smallest gap any open-weight has ever held)
• WebDev top-10 — front-end code generation that genuinely competes with proprietary leaders on React, Tailwind, and framework-specific tasks
• Code arena #8 — close behind GPT-5.2-codex, Claude Opus 4.6, and Gemini 3 Pro
• Apache 2.0 license — no MAU caps, no field-of-use restrictions, no acceptable-use policy that creates procurement friction

The procurement implication: for the first time, an enterprise can shortlist an open-weight model on capability grounds alone, deploy it under Apache 2.0, and avoid both vendor lock-in and licensing landmines. That combination did not exist 18 months ago.

For the cost-side analysis — including the 200M-token/month break-even where self-hosting GLM-4.7 stops winning on TCO — see our cross-cluster deep-dive: Open-Source LLM ROI: Why Free Costs 60% More Than Claude.

OLMo 3.1 vs Llama 4 — The Apache 2.0 vs MAU-Cap Contrast

This is the procurement comparison that matters most in 2026, because it captures the entire open-weight licensing tension in two models.

• Higher headline Elo than OLMo 3.1
• Battle-tested in production at scale across thousands of enterprise deployments
• Critical clause: the Llama Community License caps free commercial use at organizations with under 700 million monthly active users. Above that threshold, an explicit license is required from Meta.
• Practical procurement impact: for almost all enterprises, this is irrelevant. For their parent organizations (consumer brands, telcos, large platforms with end-user MAU), it becomes a compliance question. Some legal teams flag it preemptively as an unmanageable risk.

• Slightly lower Elo, still flagged Preliminary (vote count under 5,000)
• Apache 2.0 license — no MAU caps, no field-of-use restrictions
• Released by AI2 with full training data, training code, and intermediate checkpoints. The most genuinely "open" of the open-weight cohort.
• Unlike Llama 4, no procurement landmine exists at any scale.

The procurement read: if your organization's MAU is comfortably below 700M and likely to remain there, Llama 4's higher Elo wins. If MAU could plausibly cross that line — or if your legal team weighs license clauses heavily — OLMo 3.1's slightly lower Elo combined with clean Apache 2.0 wins. The 11-point Elo gap matters less than the licensing tier in most enterprise procurement contexts.

Are Mistral and DeepSeek Still Competitive in 2026?

Yes — but with caveats that didn't apply 12 months ago.

• Strong general-purpose reasoning, particularly on European languages
• Mistral Research License restricts commercial deployment without a paid agreement — a significant procurement step versus pure Apache 2.0
• Mistral Medium and Small variants under Apache 2.0 sit lower on the leaderboard but are deployment-clean
• The "research-only top tier" pattern frustrates teams expecting Apache-grade openness across the catalog

• Highest-Elo open-weight model after GLM-4.7
• DeepSeek License v2 is permissive but includes acceptable-use provisions broader than Apache 2.0
• China-origin model — triggers data-residency review in U.S. defense, EU GDPR-strict, and certain financial-services procurement
• Procurement teams in those regulated contexts often eliminate DeepSeek before capability evaluation begins, regardless of Elo

The procurement read: both remain capability-competitive, but the licensing-and-residency layer increasingly determines whether they make the shortlist. The Apache 2.0 cohort (GLM-4.7, Qwen 3.5-Coder, OLMo 3.1) holds an underweighted procurement advantage that a pure-Elo comparison hides.

Self-Hosting a Top-10 Open-Source Model — What It Actually Takes

The most expensive misreading of open-weight rankings is treating "downloadable from Hugging Face" as a synonym for "deployable in production."

The realistic stack for a top-10 open-weight model in production:

• 4-8x H100 or H200 GPUs for moderate concurrency. Quantized 4-bit deployments can reduce this to 2-4 GPUs but degrade quality measurably.
• 1.0 FTE MLOps engineer fully loaded (~$220K U.S. annually, ~$18K/month). Not a part-time devops handoff.
• Inference orchestration software — vLLM, TGI, SGLang, or proprietary equivalent. Either licensed or significant in-house engineering investment.
• Observability stack — typically Datadog or Grafana plus custom LLM-specific tracing. ~$1,500/month at moderate scale.
• Compliance audit overhead — SOC 2 attestation, security reviews, model-update regression testing.
• Capacity planning, autoscaling, and on-call rotation — non-trivial for 24/7 production.

The all-in monthly TCO floor: $18,000–$35,000 for a small-team self-hosted GLM-4.7 deployment. That math breaks even with Claude Opus 4.6 API at approximately 1.5–2 billion output tokens per month — far higher than most teams assume.

The hosted-API alternative (Together AI, Fireworks, Anyscale, OpenRouter, Groq) running open-weight models on shared infrastructure is genuinely cheaper for sub-1B-token workloads — and is the option most enterprises should actually compare against the proprietary APIs.

The Open-Weight Capability Tax — Why Per-Token Pricing Lies

The Elo gap from #1 (Claude Opus 4.6, 1504) to top-10 open-weight (GLM-4.7, 1462) understates the production gap on agentic workloads. The leaderboard measures human preference on single-turn chat. Production agentic tools (Aider, Cursor, Cline, Devin) measure something different: the rate at which a model autonomously edits multiple files, runs tests, and produces an accepted PR.

On agentic benchmarks, the gap widens:

• SWE-Bench Verified — Claude Opus 4.6 ~67%, GLM-4.7 ~52%. A 15-percentage-point gap.
• Aider Polyglot — Claude Opus 4.6 ~71%, GLM-4.7 ~58%. Comparable gap.
• Hallucinated-import rate (internal eval) — Claude ~1.8%, top open-weight ~4.1%. Doubles the retry overhead.
• Agentic-loop retry rate — Claude Opus 4.6 ~9%, top open-weight ~18%. Doubles effective token consumption on complex tasks.

The cost-per-accepted-PR translation: Claude ~$1.85, GLM-4.7 ~$3.20-3.80 once retry costs are included. The 5x per-token pricing advantage of GLM-4.7 partially or fully reverses on engineering workloads. For tolerant workloads (content generation, summarization, data extraction) where retries are cheap, GLM-4.7 still wins on net economics. For agentic engineering, Claude usually wins.

The Bottom Line — Capability Is the Easy Part

The open-weight tier has finally narrowed the proprietary capability gap. That's the headline. The procurement-grade story is less flattering.

• Capability: seven open-weight models genuinely compete with the proprietary tier on capability. The Elo gap is real but no longer disqualifying.
• Licensing: the Apache 2.0 cohort (GLM-4.7, Qwen 3.5-Coder, OLMo 3.1) is the only fully clean procurement option. Llama Community, Mistral Research, and DeepSeek License all carry asterisks that matter at enterprise scale.
• Production economics: the headline per-token cost advantage often reverses on agentic workloads where retry rates compound. Cost-per-accepted-PR is the metric that matters; cost-per-token is the metric vendors quote.
• Self-hosting: the right answer for above ~1.5B tokens monthly or for top-secret data. The wrong answer for almost everything else, where hosted-API providers (Together, Fireworks, Groq) deliver open-weight pricing without the ops overhead.

Frequently Asked Questions (FAQ)

Sources & References

• LMArena (official) — Live LLM leaderboards including the open-source ranking baseline.
• LMArena Leaderboard Changelog — Official record of model additions and methodology changes.
• arena-ai-leaderboards JSON Feed — Open mirror of official LMArena data for programmatic access.
• AI2 OLMo — Reference for OLMo 3.1 training data, Apache 2.0 license, and full reproducibility.
• Llama 4 Community License — Official license including the 700M-MAU clause.
• Aider Polyglot Leaderboard — Multi-file agentic coding benchmark for capability tax calculation.
• SWE-Bench Verified — Real GitHub issue resolution benchmark.
• Together AI Pricing — Reference hosted-API pricing for GLM, DeepSeek, Llama, Qwen, Mistral.