AivexaNewsSearch
AI news for builders and product teamsChecked every hour

[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed

Collected Oct 24, 2026

Moonshot released Kimi K3.1 Pro on October 22-23, 2026, as an upgrade to Kimi K3 built through substantially larger post-training compute rather than a new pretraining run. The model is a multimodal mixture-of-experts system Moonshot calls 3T-class: 2.8 trillion total parameters, 104 billion active per token, and a 1-million-token context window. Moonshot advertises strength in full-project coding, Blender and CAD work, financial modeling, and long-running agents.

Availability covers Kimi's website and app, Kimi Work, Kimi Code, and the API, in Standard and Highspeed variants. Moonshot claims Highspeed keeps the same intelligence while generating output 2-3x faster, and its official-account summary puts Highspeed at twice the standard price. The launch also introduced Kimi Hosted Agents, a managed harness with a 24/7 execution environment for long-running jobs and multi-agent collaboration.

The important caveat is timing: at launch the weights were not out. Moonshot promises them on November 4, 2026, and Artificial Analysis explicitly recorded the weights as unpublished. Until the license, redistribution terms, commercial-use rules, checkpoint precision and full inference requirements are known, "soon-to-be open-weight" is the accurate label rather than open source. The release also drew confusion from early rumor-driven reports and guesses about the date, including a November 5 reading, before the official November 4 date appeared; users reported the service surfacing on kimi.com before kimi.ai.

Independent results are strongest on agentic work. Artificial Analysis ranked K3.1 Pro fourth on its Intelligence Index at 55, behind Opus 5.5 at 58 and tied-group Sonnet 5.5 and Muse Wonder Full at 56. Arena placed it eighth in Code Arena: WebDev at 1,692 points, up nine places and 37 points over K3, and says it would become Arena's highest-ranked open-weight model after publication, passing DeepSeek V4.1 Flash. Arena also put it seventh in Brand & Marketing (up from ninth) and ninth in Simulations (up from twentieth), roughly matching Opus 5 Max at those settings. Artificial Analysis reported accuracy rising from 48% to 51% and hallucination rate falling from 53% to 44% on AA-Omniscience, and said the leading-open-versus-proprietary gap on its Index would narrow from 12 points to three once weights ship. A separate aggregation posted a preliminary ECI of 164.7 with a wide 152.5-169.0 bootstrap interval.

Standard first-party pricing is $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached-input tokens, down a third from K3's $3/$15/$0.30; the Chinese account quotes RMB 13, RMB 68 and RMB 1.3 per million. Per-token savings do not translate directly into cheaper completed work. On the Intelligence Index, K3.1 Pro generated about 55K output tokens per task versus K3's 48K, leaving cost near $1.98 per task. Arena measured a $1.23 median WebDev task cost, 29% below K3 and far under Opus 5 Max, GPT-6 Sol Max and Qwen3.8-Max. On the Cyber Index, roughly double the input-token consumption pushed cost to $2.81 per task, about 15% above K3. This suggests teams budgeting from list prices should measure tokens actually consumed on their own harness.

The cybersecurity story is unusual. Moonshot says it removed cyber-related data from mid-training and reinforcement learning rather than relying only on refusal behavior at inference; quoted figures show an internal cyber score dropping from 90.5% to 81% while other capabilities kept improving. Despite that, K3.1 Pro scored 64% on Artificial Analysis' Cyber Index, the highest publicly accessible result and second overall behind trusted-access GPT-6 Sol Daybreak Blue at 69%, with its index improving 23 points over K3's 41%, driven largely by CyberGym rising from 58% to 95%. Artificial Analysis observed no safety refusals on its tasks. Two mechanisms are being conflated in discussion: training-time capability reduction relative to a counterfactual model, and API refusal policy. Artificial Analysis cautions that proprietary models can hold stronger latent defensive capability than public scores show, since trusted-access GPT-6 Sol scored 32 points above its public counterpart.

Outside conventional coding the evidence is mixed. Moonshot claims 65% fewer hallucinations in financial modeling, but the baseline and protocol are unspecified, making that a vendor claim. Demonstrations include a 3D robotic-hand website split into 2,317 pieces and a Blender recreation of Singapore Changi Airport with assets, trees, scene assembly and rendering, plus native Tripo plugin support in Kimi Work. Artificial Analysis said early AA-3D preferences suggest a lower standing than in knowledge work and coding, with voting still underway. On Max Bittker's RuneScape Bench, cost was comparable to Argon with performance similar to DeepSeek V4.1 Flash and roughly Opus 5.0; a critic framed that as 52x DeepSeek's cost for similar success. One user found it weaker than Fable 5.1 and Opus 5.5 on compiler work, and a commentator noted no personal testing. These are workload anecdotes, not population results. Athletes of the benchmark debate also flagged harness dependence, notably heavy use of the Claude Code harness, and a possible provenance mismatch in the Terminal-Bench comparison chart. In practice, the model looks like a strong agentic and coding option with a genuine open-weight path ahead, not a uniform frontier leader.

Likely trade-offs follow from the architecture. 104 billion active parameters does not mean a 104-billion-parameter dense footprint: deployment must hold the whole expert pool plus runtime state, so self-hosting is a data-center undertaking, not a laptop install, and "local" claims should be read as self-hostable. As an illustrative calculation, 2.8 trillion parameters alone imply roughly 5.6 TB at 16-bit or 1.4 TB at 4-bit before quantization metadata and serving overhead, and no layer layout, routing configuration, attention design or training-compute budget has been disclosed. Serving economics remain the central business objection, with one commentator estimating 30x or more cost versus DeepSeek Flash. No Flash variant was announced. Separately, allegations of undisclosed paid promotion circulated; those concern marketing conduct and say nothing about capability.

Why it matters: developers and product teams get a near-frontier, explicitly agent-oriented model today through API, apps and a hosted agent harness, at token prices a third below K3, while the open-weight question resolves on November 4, 2026. In my reading, the practical decision is to benchmark it on real agent workloads and measure output-token consumption rather than list prices, since verbose reasoning erodes headline discounts. Teams planning to self-host should wait for license and checkpoint details, and security-sensitive groups face a novel question: a model deliberately stripped of cyber training data can still top public defensive benchmarks, so capability and policy should be evaluated separately.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

We have Fable at home... or so we are being told by paid shills