October 2026 reshuffled the open-model leaderboard in a single week. DeepSeek shipped a faster, cheaper checkpoint, Alibaba's Qwen family kept spreading across every size from laptop to data centre, and Moonshot's Kimi landed within a whisker of the best closed models on coding tests. If you are choosing an open AI model today, the question is no longer whether open models are good enough. It is which one fits your budget, your hardware, and your workload.
One warning before the numbers: most headline scores below are vendor-reported or come from a single independent index, and different test setups produce different results. Treat them as a guide, not gospel, and run your own tasks before committing a product to any model.
The three contenders at a glance
| Model | Maker | Size | Context | API price (per 1M tokens) | Weights status |
|---|---|---|---|---|---|
| DeepSeek V4.1-Flash-1017 | DeepSeek | 552B mixture-of-experts | 1M tokens | $0.15 in / $0.60 out (off-peak) | Open, MIT licence |
| Qwen3.8 (flagship tier) | Alibaba | 2.4T total, ~95B active | 1M tokens | ~$2 in / $6 out (Max API) | Open weights published; smaller sizes Apache 2.0 |
| Kimi K3.1 Pro | Moonshot AI | 2.8T total, 104B active | 1M tokens | $2 in / $10 out | API now; full weights promised by 4 November 2026 |
All three offer a 1-million-token context window, which is the real story here: long documents, whole codebases and multi-hour agent sessions are now standard fare for open models, not a premium extra.
DeepSeek V4.1-Flash-1017: the price-performance pick
DeepSeek released V4.1-Flash-1017 on 17 October 2026, describing it as a minor update to V4.1 Flash with refreshed post-training. The gains are anything but minor on agent-style work. On DeepSeek's own benchmark table, Terminal-Bench 4.0 jumped from 31.2 to 42.3, DeepSWE v1.1 rose to 77.0 (above the 74.0 DeepSeek lists for Claude Opus 5), and Humanity's Last Exam without tools rose from 36.8 to 42.6. It also leads on Automation-Bench (60.5) and Agents' Last Exam (35.4) in the same table.
The honest caveat: Claude Opus 5 still leads comfortably on several rows in that very table, including Terminal-Bench (51.8) and HLE (56.3). DeepSeek is not claiming to beat the frontier across the board, and neither should you.
Where DeepSeek genuinely wins is price. Off-peak API pricing is $0.15 per million input tokens and $0.60 per million output tokens, with peak rates exactly double. Cached input is a fraction of a cent. For high-volume agent loops, that is an order of magnitude cheaper than flagship closed models, and the MIT-licensed weights mean you can self-host without licence negotiations. If your workload is coding agents, automation or long tool-using sessions on a budget, DeepSeek is the default shortlist entry. Read our full breakdown in DeepSeek V4.1-Flash-1017: Faster AI After the Outage.
Qwen3.8: the ecosystem pick, from laptop to data centre
Qwen is the hardest family to summarise because it is not one model. Alibaba's Qwen3.8 line spans a 2.4-trillion-parameter flagship (around 95B active parameters per token) with open weights published on ModelScope, down to small models such as Qwen3.8-27B that run on a single 24GB graphics card. The flagship tier is served as Qwen3.8-Max at roughly $2 per million input tokens and $6 per million output tokens, with a 1-million-token context window.
That range is Qwen's superpower. No other family in this comparison covers the same spread: a small Apache-2.0 model you can run locally for privacy-sensitive work, a mid-size model for a single server, and a Max-class model for agent workloads. Qwen is also the most widely fine-tuned open family in the world, which means tutorials, quantised versions and community tooling are everywhere.
The trade-off is complexity. Variant names multiply fast (Flash, Max, Plus, numbered checkpoints), and the flagship API tier and the downloadable weights are not the same product, so check exactly which variant a benchmark or price quote refers to. Alibaba has also confirmed Qwen4 is in training, announced at its Apsara Conference on 22 September 2026, but with no specifications, benchmarks or release date yet — so buy for what Qwen3.8 does today, not for a promised successor.
Kimi K3.1 Pro: the closest open model to the frontier
Moonshot AI launched Kimi K3.1 Pro on 23 October 2026, and it is the headline act of this comparison. It is a 2.8-trillion-parameter model (104B active per token) with a 1-million-token context window, available now in the Kimi app, Kimi Work, Kimi Code and via API at $2 per million input tokens and $10 per million output tokens, with a faster Highspeed tier at double the price.
On Moonshot's own numbers, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. Independent evaluator Artificial Analysis gives it an Intelligence Index score of 55, behind only closed models — the highest a soon-to-be-open model has reached. That "soon-to-be" matters: at the time of writing, the full weights are promised by 4 November 2026 but not yet published, so today Kimi is an API model with an open future, not a model you can download. Moonshot also says it deliberately removed cyberattack-related training data, lowering its exploit benchmark score, a notable safety choice for an open release.
If you want near-frontier coding and agent performance without a closed provider's price tag, and you are happy to start on the API while the weights land, Kimi K3.1 Pro is the strongest single model in this comparison.
Head-to-head: which one wins what
- Best raw capability: Kimi K3.1 Pro — closest to Claude Opus 5.5 on coding benchmarks, highest independent index score of the three.
- Best value: DeepSeek V4.1-Flash-1017 — a fraction of the price of both rivals, with MIT weights you can use anywhere.
- Best for running locally: Qwen3.8 — the only family with genuinely practical small models alongside a flagship, plus the biggest fine-tuning community.
- Best for long agent work: A tie between DeepSeek (cheap enough to run long loops) and Kimi (strongest per-step reasoning).
- Best for privacy and self-hosting today: DeepSeek (MIT) and Qwen's smaller Apache-2.0 models; Kimi joins this group once its weights ship.
Also worth watching
The open race does not end with these three. Meta's Muse Spark Open (816B total, 42B active) is the leading US open-weight entry, though its custom licence is not open source, and Z.ai's GLM-5.3 scored strongly on the same independent index. Our best AI coding assistants guide covers how these models behave inside real developer tools, and Meta's Muse Wonder shows what the closed frontier looks like for comparison.
Pros and cons
DeepSeek V4.1-Flash-1017
Pros: cheapest API pricing by far; MIT licence; strong agent and coding scores.
Cons: still trails frontier models on hard reasoning; vendor benchmarks need independent confirmation.
Qwen3.8
Pros: every size from laptop to flagship; huge community and fine-tune ecosystem; 1M context.
Cons: confusing variant lineup; flagship API and downloadable weights are different products.
Kimi K3.1 Pro
Pros: near-frontier coding performance; competitive $2/$10 pricing; strong independent score.
Cons: weights not yet published at writing time; pricier than DeepSeek for high-volume use.
Who should pick what
- Indie developers and startups on a budget: DeepSeek. Start there; you can always upgrade later.
- Teams that need local or private deployment: Qwen3.8's smaller models today, DeepSeek for larger self-hosted work.
- Teams chasing maximum coding quality without going closed: Kimi K3.1 Pro on the API now, self-hosted once weights arrive.
- Enterprises with data-residency rules: Qwen's flagship weights or DeepSeek, both downloadable, with your own infrastructure.
Verdict
There is no single "best" open AI model in October 2026 — there is a best model for your constraint. If forced to pick one name for most readers, it is Kimi K3.1 Pro for capability, with DeepSeek V4.1-Flash-1017 as the value pick that most projects should actually start with. Qwen3.8 remains the safest long-term ecosystem bet. Whichever you choose, the gap between open and closed models has narrowed to single percentage points on the tests developers care about — and that gap is still closing.
FAQ
Is Kimi K3.1 Pro really open source?
Not yet at the time of writing. It is available via API, with Moonshot promising full open weights by 4 November 2026. Until those weights ship, treat it as soon-to-be-open.
Which open model is cheapest to run?
DeepSeek V4.1-Flash-1017 by a wide margin: $0.15/$0.60 per million tokens off-peak, versus $2/$10 for Kimi and roughly $2/$6 for Qwen3.8-Max.
Can I run these models on my own computer?
Qwen3.8's smaller models, yes — the 27B size fits a 24GB GPU. DeepSeek's 552B model needs server-grade hardware or a hosted API. Kimi's 2.8T model will need serious infrastructure once weights arrive.
Are vendor benchmark scores trustworthy?
Only partly. Vendors choose the tests and settings. Independent scores, like Artificial Analysis's index, help, but the best test is your own workload — run the same tasks on two models before committing.
Sources and further reading: DeepSeek's V4.1-Flash-1017 model card and changelog coverage (Digital Applied, PBX Science); Qwen3.8 flagship coverage (AI China News, Kie.ai); Kimi K3.1 Pro launch coverage (The AI Tools, GenAInews). Benchmark figures are vendor-reported unless stated as independent.
No comments:
Post a Comment