Best Open-Source LLM 2026 (Late October): DeepSeek V4.1 Flash, Meta Muse Spark Open, MiMo, GLM-5.3, MiniMax M3.1

Updated 2026-10-26: Re-ranked on the latest Artificial Analysis open-weights data. Added DeepSeek V4.1 Flash 1017 (replacing the September V4.1-Flash build), Muse Spark Open, MiniMax M3.1 Flash, Step 5 Preview, Reflection Beam and Nemotron 3.5 Super VL, and Tencent Hy4 Preview (unscored), plus a licence column for every model. Corrected Ling 3.1 Flash, whose MIT weights have been on Hugging Face since 13 October. Every model from the 5 October table stays in; Kimi K3 is now flagged because Artificial Analysis marks it deprecated. Kimi K3.1 Pro and Mistral Large 4 now sit in a "weights promised" box.

In the three weeks since the last version of this guide, Meta, MiniMax, Reflection and NVIDIA published new open weights, InclusionAI's Ling 3.1 Flash weights appeared on Hugging Face, StepFun finished the model card for Step 5 Preview, and DeepSeek shipped a point release. That release moved its Flash model from 39 to 45 on the index. The leader is still MiMo-V2.6-Pro, but it is now one index point clear of two other models, and that gap is too small to settle a model choice on its own.

This guide ranks the leading open-weights models on the index, plus a few notable mid-tier releases. A model makes the main table only if its weights and licence are confirmed on a Hugging Face model card as of 26 October 2026. Models that are only available as Ollama tags or in press coverage are left out. Every score is Artificial Analysis's Intelligence Index v4.3.2, an independent composite of ten evaluations, read from artificialanalysis.ai on the date above. Any benchmark a vendor published on its own model card is marked vendor-reported. Prices are per 1M tokens, as of October 2026.

A note on comparing scores. Artificial Analysis now displays the index as whole numbers, and so does this guide. Earlier versions of this page quoted two decimals (for example 46.32 for MiMo-V2.6-Pro); rounding them gives the same whole numbers you see here, except Llama 4 Maverick (9.99 then, 9.30 now) and Llama 4 Scout (8.08 then, 6.45 now), which Artificial Analysis has since rescored. Some older write-ups also quote the older v4.1.1 index, where scores in the 50s were common. You cannot convert between the two versions. A score that doesn't say which index version it comes from can't be checked.

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

MiMo-V2.6-Pro holds the top score at 46. DeepSeek V4.1 Flash 1017 and GLM-5.3 (Max) sit right behind it at 45. In practice, choose between these three on licence, speed and cost rather than on the score:

• MiMo-V2.6-Pro: MIT licence and $0.13 per Index task, the cheapest of the leaders. It is also slow, at about 55 output tokens/sec with roughly 3 seconds to first token on Xiaomi's API.
• DeepSeek V4.1 Flash 1017: MIT licence, $0.26 per task, about 170 tokens/sec and about 1.2s to first token. It gives up almost nothing on score and is roughly three times faster than MiMo.
• GLM-5.3 (Max): same score as DeepSeek at about 8x the cost per task ($2.01 ÷ $0.26 ≈ 7.7), under a custom glm-5.3 licence. Its smaller sibling GLM-5.3-Flash is MIT.

Speeds in this guide are Artificial Analysis medians. They move by a few percent from day to day, so we round them. The closed frontier is still well ahead. Claude Opus 5.5 at Max effort scores 58 on the same index and costs $5.98 per task. That puts the best open model 12 points behind for about 1/46th of the per-task cost ($5.98 ÷ $0.13 ≈ 46).

Open-weights ranking: October 2026

The table has one row per model, ordered by Intelligence Index v4.3.2. To make the cut, a model needed public weights on Hugging Face and a licence named on its model card. "Cost / task" is Artificial Analysis's average cost to complete one Index task. It is the fairest way to compare reasoning models, because they produce very different numbers of output tokens. Prices are first-party list rates unless marked otherwise.

Sources: Artificial Analysis leaderboard and model pages (Index v4.3.2), Hugging Face model cards and API, and the OpenRouter models API, all read 26 October 2026. Parameter counts are as stated on the model cards or by Artificial Analysis. Equal scores share a rank.

†Kimi K3: Artificial Analysis now marks Kimi K3 as deprecated and points to Kimi K3.1 Pro as its successor, so it no longer appears on the main leaderboard. Its model page still shows a score of 44 at Max effort and $2.00 per task. The weights are still downloadable at moonshotai/Kimi-K3, and it remains the strongest Moonshot model you can self-host until K3.1 Pro's weights ship.

*MiniMax M3.1 Flash: Artificial Analysis scores it 43 but still labels it proprietary, even though MiniMax pushed weights to Hugging Face on 23 October under its own community licence. That licence is non-commercial by default (details below). The model card states ~428B total / 23B active, while the repo's safetensors metadata reports a different total, so treat the size as vendor-stated.

Ling 3.1 Flash has two Artificial Analysis entries. The open-weights model scores 41. A separate "Ling 3.1 Flash 1019" entry scores 44, but Artificial Analysis marks it proprietary and we found no matching weights repository. We rank the open model only.

Not reviewed for this update. Artificial Analysis also lists two open-weights models we have not evaluated in depth: Motif 3 from Motif Technologies (MIT on Hugging Face; Artificial Analysis marks its score of 34 as an estimate) and K2 Horizon 375B A23B from the Institute of Foundation Models (Apache 2.0 on Hugging Face, score 31).

Effort tier matters. GLM-5.3, DeepSeek V4.1 Flash 1017, Kimi K3, DeepSeek V4 Pro, Muse Spark Open and Qwen3.8 27B are listed at their highest published tier (Max or Xhigh). Artificial Analysis scores GLM-5.3 at 34 on Low effort against 45 on Max. Don't compare one model's Max score with another model's default.

The top three: MiMo-V2.6-Pro, DeepSeek V4.1 Flash 1017, GLM-5.3

MiMo-V2.6-Pro: highest score, cheapest of the leaders

Xiaomi's 1.0T-total / 42B-active MoE has a 1M context window and a plain MIT licence, with ungated weights at XiaomiMiMo/MiMo-V2.6-Pro-RL. It has held the top open-weights spot since its 21 September release. At $0.435 / $0.87 per 1M tokens it costs $0.13 per Index task, the lowest of any model scoring 43 or above. Only its smaller sibling MiMo-V2.6-Flash ($0.06, score 38) is cheaper per task in the whole table.

The trade-off is latency. Artificial Analysis measures about 55 output tokens/sec and roughly 3 seconds to first token on Xiaomi's API. Along with Kimi K3 (also about 54 tok/s), that makes it the slowest of the models scoring 43 or above. Pick it for batch, agent back-ends and other high-volume jobs where throughput matters more than response time. Skip it for chat-style UX. Full detail: our MiMo-V2.6-Pro guide.

DeepSeek V4.1 Flash 1017: the best balance of speed, cost and licence

DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October 2026 as "a minor update to DeepSeek-V4.1-Flash", under MIT. DeepSeek's release note calls it "a Pareto improvement in performance and cost": higher benchmark scores and shorter outputs. On Artificial Analysis its Max tier went from 39 for the September build to 45, which made it the fastest-rising model on this list.

The API picked up the change automatically. Per DeepSeek's pricing page, the deepseek-flash ID now serves DeepSeek-V4.1-Flash-1017 at $0.30 input (cache miss) / $1.20 output in peak hours, and half that off-peak. The legacy deepseek-v4-flash names are still accepted, but they route to the 1017 model. Vendor-reported, comparing 1017 with the September build: Terminal-Bench 4.0 rose from 31.2 to 42.3 (Claude Code harness), and DeepSWE v1.1 from 74.2 to 77.0 (Mini-SWE-V2 harness).

Pick it as the default open model for interactive and agentic workloads that need a clean licence: about 170 tokens/sec, about 1.2s to first token, $0.26 per task. Deep dive: DeepSeek V4.1 Flash 1017 guide.

GLM-5.3: strong agentic model with a custom licence

Z.ai's 753B / 40B MoE ties DeepSeek at 45 (Max effort). It costs $2.01 per task, about 8x DeepSeek's $0.26 and 15x MiMo's $0.13. The full model's licence is a bespoke glm-5.3 licence (Hugging Face tags it other), unlike GLM-5.1's MIT. The smaller GLM-5.3-Flash is still MIT, scores 42, costs $0.28 per task and $0.15 / $0.50 per 1M on OpenRouter, and is the most-downloaded MoE model in the ranking table (about 6.35M Hugging Face downloads, per the HF API). See the GLM-5.3 launch guide and GLM-5.3-Flash guide.

New and newly verified open-weights models

Step 5 Preview (StepFun): 44, the best new entrant

Step 5 Preview is a 600B-total / 27B-active sparse MoE with a 1M-token window. The window covers input and output combined, and the card says it accepts text, images and video. The repo's commit history shows model files uploaded on 20 September; StepFun finished the model card between 13 and 15 October. Artificial Analysis scores it 44 at $1.03 per task and measures about 210 tokens/sec, the fastest of the models scoring 43 or above. It costs $1.00 / $2.70 per 1M on OpenRouter. The licence is StepFun Model License 1.0, which the card describes as "an Apache-2.0-Style license, with additional attribution and commercial-license conditions specific to model use, modification and distribution." The licence text sets three conditions:

• "Powered by StepFun AI" attribution once a product passes 100M monthly active users or $20M in revenue in a calendar month.
• A separate commercial licence for device makers shipping the model as a device feature.
• A separate commercial licence for anyone running a Model-as-a-Service (MaaS) business while their group's combined revenue from all sources exceeds $20M over any 12 months.

Internal use that doesn't make the model or its outputs available to third parties is exempt from the separate commercial-licence requirement (the device-maker and MaaS conditions). The attribution rule still applies. Vendor-reported: Terminal-Bench v4 33.3 and DeepSWE v1.1 67.7 (SWE-agent harness), at the High reasoning setting.

MiniMax M3.1 Flash: cheapest strong API, restrictive licence

At $0.15 / $0.60 per 1M tokens and $0.22 per Index task, MiniMax M3.1 Flash scores 43 on Artificial Analysis. Only MiMo-V2.6-Pro beats it on cost per task among models scoring 43 or above. The licence is the catch. The MiniMax Community License grants rights "for non-commercial purposes". Commercial use requires two things:

• Displaying "Built with MiniMax M3.1-Flash".
• Either a one-time notice to MiniMax, or prior written authorisation if the product makes more than $20M in yearly revenue.

The licence's definition of "Commercial Use" explicitly includes serving fine-tuned derivatives. Vendor-reported: Terminal Bench 4.0 38.1 (Codex 0.149.1 harness) and DeepSWE v1.1 73.7 (mini-SWE-agent 2.4.5, 512K context). See our MiniMax M3.1 Flash guide.

Ling 3.1 Flash (InclusionAI): MIT, and the cheapest hosted option

Correction. An earlier version of this guide said InclusionAI had no Ling 3.1 repository on Hugging Face and listed the licence as unknown. That is no longer true. inclusionAI/Ling-3.1-flash was created on 13 October 2026, the weights were uploaded the same day, and it is licensed MIT.

It is a 560B-total / 25B-active hybrid-linear MoE (512 routed experts, 8 active per token) with a 1M window. Artificial Analysis scores the open model 41 at $1.01 per task using InclusionAI's $0.30 / $0.90 API, and measures it at about 210 tokens/sec. The cheapest route we found is OpenRouter at $0.075 / $0.225 per 1M tokens, with a 262K context on that provider. The card recommends SGLang, with example configs for 8×H200 or 4×GB300.

Muse Spark Open (Meta): open weights from Meta again, gated and attributed

Meta Superintelligence Labs released Muse Spark Open on 15 October. It is an 816B-total / 42B-active MoE with a 1.8B-parameter Perception Encoder. It takes text, image, video and audio input, outputs text, and has a 1M context. Artificial Analysis scores it 39 (Xhigh), matching the figure Meta prints on its own card, at $0.84 per task and $1.25 / $4.25 per 1M tokens.

Where it stands out is tool use. Meta reports MCP-Atlas 89.2 and OSWorld Verified 80.4 (vendor-reported, Xhigh). It is also the only model here that Meta says is "optimized to run on two Mac Studios with M5 Ultra using MLX." The Muse Model License has three conditions to note:

• Downloads are manually gated.
• "Built with Muse Spark Open" must be displayed in products that use the model.
• MaaS providers above $20M in revenue must sign a separate agreement with Meta.

More in our Muse Spark Open guide, and for the background on Meta's shift away from Llama, see Meta's return to open weights.

Also verified: Reflection Beam 1.0, Nemotron 3.5 Super VL, Tencent Hy4 Preview

• Reflection Beam 1.0: 501B / 23B MoE, text-only, 1M context, Apache 2.0, released in BF16, FP8 and NVFP4. Artificial Analysis scores it 30 at $0.14 per task and measures about 365 tokens/sec. Among Apache-2.0 models, only Qwen3.8 27B (34) and the unreviewed K2 Horizon 375B A23B (31) score higher.
• NVIDIA Nemotron 3.5 Super VL: 120B / 12B, combining Mamba-2, MoE and attention layers with a vision encoder, 1M context, OpenMDW-1.1 licence. The repo was created on 12 October. It scores 26, but at about 450 tokens/sec it is by far the fastest model in the ranking table, and it is small enough for a single high-memory node.
• Tencent Hy4 Preview: 770B / 49B MoE, 1M context, Apache 2.0. Not an October release: the repo dates from 27 August, but this is the first version of the guide to verify it. The weights and licence are confirmed, but Artificial Analysis has not published a Hy4 score yet (its leaderboard shows the older Hy3), so it is unranked.

Also ranked: Kimi K3, Qwen3.8, MiMo-V2.6-Flash, DeepSeek V4 Pro

These models carried over from the 5 October version. None has had a new release since.

• Kimi K3 (44, Max): at 2.8T total / 104B active, the largest model in the table and the most expensive per token ($3.00 / $15.00). The Kimi K3 License requires a separate agreement with Moonshot for MaaS businesses whose group revenue exceeds $20M over 12 months, and "Kimi K3" on the UI above 100M monthly active users or $20M in monthly revenue. Internal use is exempt from both. See our Kimi K3 guide.
• Qwen3.8-2.4T-A95B (40) and Qwen3.8-Flash-Next (40): Alibaba's best open models. Neither is Apache 2.0. The 2.4T model's Qwen3.8-Max License requires a separate licence for MaaS or "AI Work Assistant" businesses above $50M in 12-month group revenue. Flash-Next's Qwen Community License 1.0 requires one for any MaaS or AI Work Assistant business, with no revenue floor. At $0.37 per task, Flash-Next is far cheaper to run than the 2.4T model ($2.16). See our Qwen 3.8 guide.
• MiMo-V2.6-Flash (38): MIT, 309B / 15B, and the cheapest model in the table at $0.06 per Index task ($0.14 / $0.28 per 1M).
• DeepSeek V4 Pro 0813 (36, Max): MIT, 1.6T / 49B. It now scores 9 points below DeepSeek's own V4.1 Flash 1017 and costs more than twice as much per task ($0.67 vs $0.26).

We'll re-rank this guide after 4 November, once both deadlines have passed.

Vendor-reported coding benchmarks (read with care)

These figures come from each lab's own model card or launch post. They are not comparable across vendors, because each lab uses a different harness, scaffold and attempt budget. Use them to see how a lab says its own model improved, not to rank models against each other. Rank with the Artificial Analysis column above.

Muse Spark Open reports Terminal-Bench 2.1 (76.4) instead of 4.0, so it isn't in this table. Ling's numbers come from InclusionAI's own LingCode harness, which it says "delivers significant performance gains on some coding agent tasks". Both are good examples of why vendor numbers need checking before you put them side by side.

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

The model families named in this post's URL are scored on Intelligence Index v4.3.2, and none of them is competitive at the top. Scores below were re-read on 26 October. None of these models has had a new release since the last version of this guide.

Meta's open-weights story has moved from Llama to Muse Spark Open, which scores 39 against Maverick's 9. Llama 4 Scout's 10M-token window is still the largest of any open model, but that is its only advantage. Gemma 4 is still the pick when the model has to run on a laptop under Apache 2.0. Artificial Analysis now marks Mistral Large 3 as deprecated; for Mistral, the question is whether Large 4's weights actually arrive.

How do the licences compare for commercial use?

Eight of the sixteen ranked models use something other than MIT or Apache 2.0. Read the terms before you commit, especially if you resell API access or ship on hardware.

If your legal team requires an OSI-approved licence, the shortlist is: MiMo-V2.6-Pro (46), DeepSeek V4.1 Flash 1017 (45), GLM-5.3-Flash (42), Ling 3.1 Flash (41), MiMo-V2.6-Flash (38), DeepSeek V4 Pro 0813 (36), Qwen3.8 27B (34) and Reflection Beam (30). Because the #1 model is MIT-licensed, requiring an OSI licence costs you nothing in index score.

Worked cost example: 20M input + 4M output tokens a day

Here is a mid-sized agent back-end at list prices, as of October 2026. Formula: (input M × input price) + (output M × output price).

Token prices only tell part of the story. Reasoning models differ a lot in how many output tokens they spend on the same task. That is why MiMo-V2.6-Pro costs more than DeepSeek per token here but less per Index task ($0.13 vs $0.26). Ling is the opposite case: cheapest per token on OpenRouter, but $1.01 per task at InclusionAI's own rates. Before you commit, benchmark on your own prompts and measure the actual output token counts. DeepSeek's off-peak rates are half the peak figures above, which matters for batch jobs you can schedule.

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

• Best all-round, clean licence: DeepSeek V4.1 Flash 1017. Scores 45, MIT, about 170 tok/s, $0.26 per task.
• Highest score, latency-tolerant: MiMo-V2.6-Pro. Scores 46, MIT, $0.13 per task.
• Lowest cost per task: MiMo-V2.6-Flash. Scores 38, MIT, $0.06 per task.
• Cheapest strong hosted API: MiniMax M3.1 Flash or DeepSeek V4.1 Flash 1017 (both $0.15 / $0.60 per million input / output tokens as of October 2026; DeepSeek via OpenRouter or off-peak first-party), or Ling 3.1 Flash on OpenRouter.
• Fastest of the top tier: Step 5 Preview at about 210 tok/s, if its licence suits you.
• Agentic tool use and computer use: Muse Spark Open, going by Meta's vendor-reported MCP-Atlas and OSWorld numbers. Check them on your own tasks.
• Multimodal input on open weights: MiMo-V2.6-Pro or Muse Spark Open (text, image, video and audio). MiniMax M3.1 Flash and Step 5 Preview take text, images and video; GLM-5.3-Flash, DeepSeek V4.1 Flash and Nemotron 3.5 Super VL, among others, also accept images.
• Apache 2.0 only: Qwen3.8 27B (34), Reflection Beam (30) or Tencent Hy4 Preview (unscored). K2 Horizon 375B A23B (31) is also Apache 2.0, but we haven't reviewed it.
• Raw throughput on modest hardware: Nemotron 3.5 Super VL. 120B / 12B, about 450 tok/s on hosted endpoints.
• On-device, laptop or air-gapped: Gemma 4 26B-A4B or Qwen3.8 27B.
• Waiting on a frontier-level open model: watch Kimi K3.1 Pro's 4 November weights deadline.

How much hardware does self-hosting need?

Every top-ranked model is a large MoE. Active parameters drive per-token compute, but you still need memory for all the weights. As a floor, weights alone take about 1 byte per parameter at 8-bit and 2 bytes at BF16, before KV cache. Requirements stated on the model cards, plus that arithmetic for the leaders:

• MiMo-V2.6-Pro (1.0T): roughly 1TB of weights at 8-bit and 2TB at BF16, so plan on multiple nodes at BF16.
• GLM-5.3 (753B): roughly 0.75TB at 8-bit and 1.5TB at BF16.
• DeepSeek V4.1 Flash 1017 (552B): roughly 0.55TB at 8-bit. The repo ships reference inference code, and community GGUF and EXL3 quantizations exist (e.g. unsloth/DeepSeek-V4.1-Flash-1017-GGUF).
• Step 5 Preview (600B): the card's vLLM and SGLang examples run the FP8 checkpoint on one 8×H200 node (or 4×B300) and the BF16 checkpoint on two 8×H200 nodes.
• Kimi K3 (2.8T): the heaviest model in the table, roughly 2.8TB at 8-bit.
• Muse Spark Open: "at least 700GB of VRAM for NVFP4 and 2TB for BF16", or two M5 Ultra Mac Studios via MLX.
• Ling 3.1 Flash: SGLang examples on 8×H200 (TP8 + EP8) or 4×GB300.
• Reflection Beam: ships official FP8 and NVFP4 checkpoints, which reduces the memory needed compared with BF16.
• Nemotron 3.5 Super VL: 120B total with an official NVFP4 build, the most practical single-node option among the newer models.

When does self-hosting pay off? On GPU rental alone, ignoring ops time, and as of October 2026, on-demand H100 80GB GPUs start at $1.99/hr on RunPod (Lambda lists H100 SXM at $4.29/hr), and H200s at $3.59/hr on RunPod Community Cloud. Two worked break-evens against the hosted prices above:

• Single GPU, Qwen3.8 27B: one H100 costs $1.99 × 24 = $47.76/day. Hosted on OpenRouter, the 20M + 4M mix costs 20 × $0.20 + 4 × $2.50 = $14.00/day, or $14.00 ÷ 24 ≈ $0.58 per 1M tokens. Break-even is $47.76 ÷ $0.58 ≈ 82M tokens/day, about 950 tokens/sec sustained around the clock.
• One 8×H200 node, Step 5 Preview (FP8): 8 × $3.59 × 24 = $689.28/day. Hosted on OpenRouter, the same mix costs $30.80/day, or about $1.28 per 1M tokens. Break-even is $689.28 ÷ $1.28 ≈ 537M tokens/day, about 6,200 tokens/sec sustained.

Both assume your hardware can actually sustain that throughput, and neither counts engineering and on-call time. Below those volumes, hosted APIs win. Self-hosting is worth it when data residency, fine-tuning or very large steady volumes require it. For choosing a serving stack, see our vLLM vs Ollama vs LM Studio production comparison.

FAQ

What is the best open-source LLM right now?

As of 26 October 2026, MiMo-V2.6-Pro leads Artificial Analysis's Intelligence Index v4.3.2 among open-weights models at 46. DeepSeek V4.1 Flash 1017 and GLM-5.3 (Max) are tied at 45. MiMo and DeepSeek are MIT-licensed; GLM-5.3 uses a custom licence.

Is Kimi K3.1 Pro open source?

Not yet. Moonshot says "the full model weights will be released by Nov 4," but no weights or licence were public on 26 October. At 55 on the index it would lead the open field by 9 points if it ships. Until then, Kimi K3 (44) is the strongest Moonshot model with downloadable weights.

Are Mistral Large 4's weights available?

Not as of 26 October. Mistral launched a preview on 6 October and wrote that it would "release the weights by the end of the month." Until then, Mistral Large 3 (Apache 2.0) is Mistral's latest open-weights large model.

Can I use MiniMax M3.1 Flash commercially?

Only on MiniMax's terms. The MiniMax Community License is non-commercial by default. Commercial use requires "Built with MiniMax M3.1-Flash" attribution and a notice to MiniMax, or prior written authorisation above $20M in yearly revenue. Using MiniMax's paid API is a separate question from self-hosting the weights.

Is Ling 3.1 Flash open weights?

Yes. InclusionAI published the weights on Hugging Face on 13 October 2026 under MIT. Artificial Analysis also lists a "Ling 3.1 Flash 1019" checkpoint that scores 44, but labels it proprietary, so we don't rank it.

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

DeepSeek, clearly. DeepSeek V4.1 Flash 1017 scores 45. Qwen's best open models, Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next, score 40 each, but under licences that restrict MaaS businesses; the Apache-2.0 Qwen3.8 27B scores 34. Llama 4 Maverick and Mistral Large 3 both score 9. Meta's current open model is Muse Spark Open at 39, not Llama.

Can I run any of these on a single GPU?

Gemma 4 26B-A4B and Qwen3.8 27B are the realistic single-GPU picks. At BF16, Qwen3.8 27B's weights take about 55GB, which fits one 80GB H100; 4-bit quantisation brings that to roughly 14GB. GLM-5.3-Flash (320B) and MiMo-V2.6-Flash (309B) need a multi-GPU node even when quantised: at 4-bit their weights alone are roughly 155–160GB. MiMo-V2.6-Pro, GLM-5.3, DeepSeek V4 Pro and Kimi K3 are multi-node models at BF16.

How far behind closed models are open-weights models?

Twelve index points: 46 for MiMo-V2.6-Pro against 58 for Claude Opus 5.5 (Max). On cost per task, the open leader is about 46x cheaper ($0.13 vs $5.98).

Should I trust vendor-reported benchmarks?

Treat them as directional. Artificial Analysis runs every model on the same index version, so its scores are comparable with each other. Vendor Terminal-Bench and DeepSWE numbers use different harnesses, and some cards report different benchmark versions (Terminal-Bench 2.1 vs 4.0).

How often does this ranking change?

Often. Between 5 and 26 October, DeepSeek's Flash model gained six index points from one point release, and Step 5 Preview and MiniMax M3.1 Flash entered the top 10. Re-check against the current index before you lock in a model, and build a model router and evaluation harness instead of hard-coding one model.

Sources

• Artificial Analysis: model leaderboard (Intelligence Index v4.3.2)
• Artificial Analysis: open-weights models
• Artificial Analysis: MiMo-V2.6-Pro
• Artificial Analysis: DeepSeek V4.1 Flash 1017
• Artificial Analysis: GLM-5.3
• Artificial Analysis: Kimi K3 (deprecated)
• Artificial Analysis: MiniMax-M3.1-Flash
• Artificial Analysis: Ling 3.1 Flash
• Artificial Analysis: Muse Spark Open
• Artificial Analysis: DeepSeek V4 Pro 0813
• Artificial Analysis: Mistral Large 3 and Llama 4 Maverick
• Artificial Analysis: Claude Opus 5.5
• Hugging Face: Muse-Spark-Open-816B-A42B
• Hugging Face: DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813
• Hugging Face: MiniMax-M3.1-Flash and its licence
• Hugging Face: Step-5-Preview and StepFun Model License 1.0
• Hugging Face: Ling-3.1-flash
• Hugging Face: MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL
• Hugging Face: GLM-5.3 and GLM-5.3-Flash
• Hugging Face: Kimi-K3 and Kimi K3 License
• Hugging Face: Qwen3.8-2.4T-A95B, Qwen3.8-Flash-Next and Qwen3.8-27B
• Hugging Face: Reflection Beam-1.0-501B-A23B
• Hugging Face: NVIDIA Nemotron 3.5 Super VL
• Hugging Face: Tencent Hy4-preview
• Hugging Face: Motif-3 and IFM K2-Horizon-375B-A23B
• DeepSeek: V4.1-Flash-1017 release note
• DeepSeek: API pricing
• Moonshot AI: Kimi K3.1 Pro launch post
• Mistral AI: Mistral Large 4 preview
• OpenRouter: models API (hosted prices)
• RunPod: GPU pricing and Lambda: GPU pricing

If your team is putting one of these models into production, whether that's serving, evals or a model router, Codersera can help you hire vetted remote AI developers who have run open-weight LLM infrastructure before.