Best Open-Source LLM 2026: MiMo, GLM-5.3, Kimi, DeepSeek

The open-weights frontier turns over roughly every six weeks, and the honest answer to "what's the best open source LLM?" in October 2026 shares almost no names with the answer from May. Kimi K2.6, GLM-5.1, MiMo-V2.5-Pro and DeepSeek V4-Flash — the models that topped this list earlier in the year — have all been superseded by their own successors. Llama 4, Qwen 3.5, Gemma 4 and Mistral Large 3 are now measured, and they sit far below the leaders.

This guide ranks the current leaders on the Artificial Analysis Intelligence Index v4.3.2 — a neutral composite of ten independent evaluations (AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR). Every score below was read from Artificial Analysis on 5 October 2026 and is stamped with its index version. Vendor-reported numbers are labelled as such.

Read this before comparing against another article. Artificial Analysis rebased the Intelligence Index from v4.1.1 to v4.3.2. The benchmark basket changed, so every score moved — downward, across the whole board — without any model getting worse. There is no conversion factor. A v4.1.1 figure (the 54s, 52s and 57s that circulated through mid-2026) cannot be compared with a v4.3.2 figure, and a score quoted without its index version is unverifiable. If a number on another page is not stamped, assume you cannot use it.

The rebase also explains why the order changed, not just the magnitudes. v4.3.2 added Terminal-Bench 4.0 and AutomationBench-AA to the basket, taking it to ten evaluations. Both measure long-horizon agentic execution — driving a shell, recovering from failed steps, finishing a multi-stage task — so agentic capability now carries far more weight than it did under v4.1.1. Models tuned for chat and single-turn reasoning fell; models built to operate tools rose. That single change is most of the reason the open-weights leaderboard reshuffled rather than merely rescaled.

What is the best open-source LLM in October 2026?

MiMo-V2.6-Pro, from Xiaomi. On Intelligence Index v4.3.2 it scores 46.32, ahead of every other open-weights model Artificial Analysis has measured, and it does so under a plain MIT licence with ungated weights on Hugging Face (XiaomiMiMo/MiMo-V2.6-Pro-RL, ~86k downloads in the last 30 days per the Hugging Face API).

The gap to the closed frontier is the number worth internalising. Claude Opus 5.5 (Max) scores 57.62 on the same index — 11.3 index points ahead of the best open model. That is a wider gap than the open-weights narrative claimed at mid-year, and the rebase is part of why: v4.3.2 leans harder on long-horizon agentic work, which is exactly where open models still lose. On Terminal-Bench 4.0, the agentic shell benchmark inside the index, Opus 5.5 scores 59.6% against GLM-5.3's open-weights-leading 41.9%.

But "best" is workload-dependent, and on cost it is not close at all. MiMo-V2.6-Pro completes an average Index task for $0.13 against Opus 5.5's $5.98 — about 45x cheaper per completed task for 80% of the intelligence. That trade is the whole reason this category matters.

Which open-weight LLMs rank highest on the v4.3.2 index?

One row per model, ordered by Intelligence Index v4.3.2. "Cost / task" is Artificial Analysis's average cost to complete one Index task — the right unit for cross-model comparison, because AA runs different task counts per model and whole-suite totals are therefore not comparable. Prices are first-party API list rates per 1M tokens.

Source: Artificial Analysis model pages, Intelligence Index v4.3.2, read 5 October 2026.

Two things to check before you quote any of these numbers elsewhere. First, effort tier is stated above wherever a model publishes more than one, because the gaps are enormous: GLM-5.3 scores 44.78 at Max effort and 34.30 at Low, and Kimi K3 scores 43.59 at Max and 30.07 at Low. A table that mixes tiers is not a comparison. Second — and this trips up most write-ups — Artificial Analysis's unqualified model label is not a consistent tier. The bare "GLM-5.3" and "Kimi K3" records are their Low variants; the bare "GPT-5.5" record is its Xhigh variant. Resolve the tier explicitly rather than assuming the default is the top of the ladder. MiMo-V2.6-Pro publishes a single variant, so its 46.32 needs no qualifier.

One caveat on row 5: Artificial Analysis classifies Ling 3.1 Flash as open-weights but publishes no licence name, and Ant Group's InclusionAI had no Ling-3.1 repository on Hugging Face when we checked, so treat its weights availability as unconfirmed.

Why is MiMo-V2.6-Pro the top open-weights model?

Xiaomi's MiMo-V2.6-Pro is a 1,020B-total / 42B-active Mixture-of-Experts model with a 1M-token context window, MIT-licensed, serving at $0.435 input / $0.87 output per 1M tokens. It leads the open field on Intelligence Index v4.3.2 at 46.32 and backs that up on the components that matter for engineering work: GDPval-AA 1686 (real-world work tasks), Terminal-Bench 4.0 34.8%, SciCode 60.9%, Humanity's Last Exam 49.4% and AA-LCR 86.3% on long-context reasoning.

The economics are the headline. At $0.13 per Index task it is the cheapest model in the entire top tier — cheaper per task than GLM-5.3 (15x), Kimi K3 (15x) and Qwen3.8-2.4T (16x), all of which it outscores. No other model on this list wins on capability and cost simultaneously.

Two caveats, both real. MiMo is slow: Artificial Analysis measures a median 41 output tokens/sec and a 4.2s time to first chunk, which is the worst latency profile of any model in the top five. And ecosystem maturity lags DeepSeek and Qwen — fewer hosted providers, less battle-tested serving tooling.

Pick MiMo-V2.6-Pro if you want the highest open-weights capability under a genuinely clean MIT licence and your workload is batch or throughput-shaped rather than latency-sensitive. Skip it if you are serving interactive UX, where 41 tok/s will be felt. Full detail: our MiMo-V2.6-Pro guide.

Where do GLM-5.3 and Kimi K3 land?

GLM-5.3 (Z.ai) is the agentic pick. A 753B-total / 40B-active MoE with a 1M context, it scores 44.78 at Max effort and posts the highest Terminal-Bench 4.0 score of any open model at 41.9% — closer to the closed frontier's 59.6% than anything else with downloadable weights. Its GDPval-AA of 1653 is likewise near the top of the open set. The cost is literal: $2.01 per Index task, 15x MiMo's.

The licence changed, and this matters if you chose GLM-5.1 for its licensing. GLM-5.1 shipped under plain MIT; GLM-5.3 ships under a bespoke "GLM-5.3 Licence" (Hugging Face categorises it as other). The smaller GLM-5.3-Flash is still MIT — and at 41.81 for $0.25 per task and $0.15/$0.50 per 1M tokens, it is the best option on this list for teams whose legal review stops at "is it MIT or Apache?". It is also the most-downloaded model here by a wide margin (~5.7M Hugging Face downloads in 30 days).

Kimi K3 (Moonshot) is the long-context and knowledge pick. At 2,800B total / 104B active it is the largest model on this list, with a 1,048,576-token window. It scores 43.59 at Max effort, and leads the open field on two specific components: AA-LCR 88.7% (long-context reasoning) and AA-Omniscience +19.7 — the only open model that approaches GPT-5.5's +20.5 on factual reliability. It is weak where GLM-5.3 is strong: Terminal-Bench 4.0 of just 12.6%. K3 also ships under a bespoke Kimi K3 Licence, not the Modified MIT that covered K2.6, and at $3/$15 per 1M tokens it is the most expensive open model here.

Deep dives: GLM-5.3 launch guide, GLM-5.3-Flash guide, Kimi K3 complete guide.

Is DeepSeek still competitive in October 2026?

Yes — but not with the models most articles still name. DeepSeek V4-Flash is retired. Per DeepSeek's pricing page, the current model name is deepseek-flash, and the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names are "still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model."

DeepSeek V4.1-Flash (released 10 September 2026) is the model to evaluate. It is a 552B-parameter MoE using a new causal encoder–decoder design — DeepSeek's release notes state "just 8B active parameters for input, 16B for output" — with native multimodal input, a 1M context, and MIT weights at deepseek-ai/DeepSeek-V4.1-Flash. It scores 39.46 at Max effort for $0.27 per Index task.

The internal upset is the interesting part: V4.1-Flash outscores DeepSeek's own flagship V4 Pro (36.00) by 3.5 index points, beats it on Terminal-Bench 4.0 (26.8% vs 14.1%) and on GDPval-AA (1600 vs 1455), runs roughly twice as fast (213 vs 110 median output tok/s), and costs a quarter as much per task. V4 Pro remains on the API — DeepSeek's changelog states that "in response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026" — but it is no longer the right default inside its own family. Pro's one clear edge is factual reliability (AA-Omniscience +0.8 against Flash's −5.3).

See our DeepSeek V4.1-Flash complete guide for the full breakdown.

What happened to Llama 4, Qwen 3.5, Gemma 4 and Mistral?

All four are now scored on Intelligence Index v4.3.2, and none is competitive. This is the single biggest correction to earlier versions of this guide, which listed Llama 4 and Gemma 4 as unscored and therefore gave them the benefit of the doubt.

The practical reading: Llama 4's only remaining claim is Scout's 10M-token context window, which is still the largest of any open model and roughly 10x the 1M windows above it. If RAG-on-everything in a single call is your requirement, Scout is defensible. For anything where answer quality matters, it is 38 index points behind the open leader. Gemma 4 keeps its category of one — a model that genuinely runs on a laptop, under Apache 2.0, with the 26B-A4B variant the stronger of the two. Qwen 3.5 is simply superseded: Qwen3.8-Flash-Next scores 39.82 against Qwen3.5-397B's 18.42 while activating 6B parameters instead of 17B. Mistral remains an Apache-2.0 and EU-jurisdiction answer, not a capability answer.

How do the licences compare for commercial use?

Licence cleanliness moved backwards in 2026 at the very top of the board. Two of the three leading models adopted bespoke licences where their predecessors were MIT or Modified MIT.

If your legal team mandates an OSI-approved licence, your frontier shortlist is now four names: MiMo-V2.6-Pro (46.32, MIT), GLM-5.3-Flash (41.81, MIT), DeepSeek V4.1-Flash (39.46, MIT) and Qwen3.8-27B (33.70, Apache 2.0). That is a narrower list than mid-2026 offered, but MiMo-V2.6-Pro sitting at the top of it means the OSI constraint now costs you nothing at all in capability.

Which open-source LLM should you pick for your workload?

• Best all-round open model? MiMo-V2.6-Pro — 46.32 on Index v4.3.2, MIT, and the cheapest per task in the top tier.
• Best agentic / shell-and-tool work? GLM-5.3 at Max effort — Terminal-Bench 4.0 of 41.9%, the open-weights high.
• Best factual reliability? Kimi K3 — AA-Omniscience +19.7, the only open model near the closed frontier on this axis.
• Best long-context reasoning? Kimi K3 (AA-LCR 88.7%, 1.05M window). For raw window size only, Llama 4 Scout's 10M.
• Best capability per dollar? MiMo-V2.6-Flash at $0.06 per Index task for a 37.88 score, or MiMo-V2.6-Pro at $0.13 for 46.32.
• Cleanest licence with no capability sacrifice? MiMo-V2.6-Pro (MIT) or GLM-5.3-Flash (MIT).
• Multimodal input on open weights? DeepSeek V4.1-Flash — native vision, MIT, 1M context.
• Fastest interactive serving? DeepSeek V4.1-Flash at ~213 median output tok/s and ~1.05s to first chunk.
• On-device / laptop / air-gapped? Gemma 4 26B-A4B or Qwen3.8-27B.
• EU-jurisdiction requirement? Mistral Large 3, accepting the capability cost.

For model-specific depth see our Kimi K3 guide, Qwen 3.8 guide, and AI Coding Agents Complete Guide (2026).

How much does it cost to self-host these models?

The threshold math still holds, and the cheaper hosted rates at the small end have pushed it up, not down. A single H100 runs ~$2–4/hr on-demand; H200 ~$3–5/hr. Self-hosting beats a hosted API once you cross roughly:

• ~5M tokens/day sustained for a single-host model — Qwen3.8-27B, Gemma 4, or a quantised MiMo-V2.6-Flash / GLM-5.3-Flash.
• ~30–50M tokens/day for a multi-host frontier MoE — MiMo-V2.6-Pro, GLM-5.3, DeepSeek V4.1-Flash, Kimi K3. Note that active-parameter count, not total, drives serving cost: DeepSeek V4.1-Flash activates 16B and GLM-5.3-Flash 18B, so both are far cheaper to serve than their totals suggest. Kimi K3's 104B active is the expensive end.

Below those thresholds, hosted APIs win on total cost once you price in ops time — and at $0.15/$0.50 for GLM-5.3-Flash, the bar for "cheaper to run it ourselves" is higher than it was a year ago. For the tradeoffs in detail, see our self-hosting LLMs guide.

FAQ

What is the best open-source LLM right now?

Xiaomi's MiMo-V2.6-Pro, at 46.32 on Artificial Analysis's Intelligence Index v4.3.2 (read 5 October 2026) — the highest score of any open-weights model, under a plain MIT licence, at $0.13 per Index task. GLM-5.3 (44.78) and Kimi K3 (43.59) are next.

Llama 4 vs DeepSeek vs Qwen vs Mistral — which is best in 2026?

DeepSeek, clearly. On Intelligence Index v4.3.2 the best current model from each family is DeepSeek V4.1-Flash at 39.46, Qwen3.8-2.4T-A95B at 39.89 (but under a revenue-gated licence; the Apache-2.0 Qwen3.8-27B scores 33.70), Qwen 3.5's best at 18.42, Mistral Large 3 at 9.27, and Llama 4 Maverick at 9.99. Llama 4's only remaining advantage is Scout's 10M-token context. None of the four leads the open field overall — MiMo-V2.6-Pro does, at 46.32.

How far behind closed models are open-weights models?

11.3 index points on v4.3.2: 46.32 for the best open model against 57.62 for Claude Opus 5.5 (Max). The gap is widest on agentic work — Terminal-Bench 4.0 of 41.9% for the best open model versus 59.6% for Opus 5.5 — and narrowest on long-context reasoning, where Kimi K3's AA-LCR of 88.7% actually exceeds Opus 5.5's 84.7%.

Why are these scores lower than the ones I saw earlier in 2026?

Artificial Analysis rebased the Intelligence Index from v4.1.1 to v4.3.2, changing the underlying benchmark basket. Every score moved down without any model changing. There is no conversion factor between the two versions, so a v4.1.1 number (the 50s and 57s widely quoted mid-year) cannot be compared with a v4.3.2 number. Always check which version a score is stamped with.

Is DeepSeek V4-Flash still available?

No. DeepSeek's pricing page states that V4-Flash and V4-Flash-Vision-Exp are retired and that requests to those legacy model names are served by DeepSeek-V4.1-Flash and billed at the Flash price. The current model name is deepseek-flash. V4 Pro is still offered as a separate model.

Which open LLM has the cleanest licence for commercial use?

MiMo-V2.6-Pro, GLM-5.3-Flash, DeepSeek V4.1-Flash and DeepSeek V4 Pro are all plain MIT. Qwen3.8-27B, Gemma 4 and Mistral Large 3 are Apache 2.0. The full GLM-5.3, Kimi K3 and Qwen3.8-2.4T models moved to bespoke licences in 2026 — GLM-5.1 and Kimi K2.6 were MIT and Modified MIT respectively, so this is a regression worth checking if you standardised on either.

Can I run any of these on a single GPU?

Gemma 4 26B-A4B and Qwen3.8-27B are the realistic single-GPU picks. GLM-5.3-Flash (18B active) and MiMo-V2.6-Flash (15B active) are single-host feasible with quantisation. MiMo-V2.6-Pro, GLM-5.3, DeepSeek V4 Pro and Kimi K3 are multi-host frontier MoEs; Kimi K3's 2.8T total and 104B active is the heaviest on this list.

Should I trust vendor-reported benchmarks?

Treat them as directional. Every score in this guide comes from Artificial Analysis, which runs its own evaluations on a stated index version, so the numbers are comparable to each other. Vendor-published SWE-Bench and Terminal-Bench figures from model cards are not comparable across labs, because harness, scaffold and attempt budget differ.

How often does this ranking change?

Roughly every six weeks the top three change. Between May and October 2026 the open-weights leader turned over three times: Kimi K2.6 → MiMo-V2.5-Pro → MiMo-V2.6-Pro. Re-validate your shortlist against the current Intelligence Index version each quarter, and check the version stamp rather than the number.

Need engineers who can actually run these models in production?

Picking a model is a one-week decision; operating a self-hosted LLM stack is a year-long engineering investment — multi-node serving, FP4/FP8 quantization, expert-routing tuning, eval pipelines against your real workload, and 24/7 on-call. The six-week leader turnover above is itself the argument for building a router and an eval harness rather than standardising on one model. If you're hiring vetted remote developers experienced with open-weight LLM infrastructure, inference optimization, and agentic coding harnesses, codersera.com/hire matches you with engineers who have shipped exactly this before, with a risk-free trial so you can validate fit before committing.