Moonshot 7/12 Kimi K2.7-0913: re-tag of K2 weights emphasizing instruction following + agent scaffolds

2 sources

Moonshot AI on 2026-07-12 announces Kimi K2.7-0913 — a re-tag of K2-line weights (architecture unchanged: 1.1T total / 32B activated MoE, Apache-2.0 with the K2 acceptable-use addendum) emphasizing long-form instruction adherence + agent scaffolds (browser / shell / file-system tool handlers) + reduced refusal on benign long-context reasoning. Self-reported gains vs K2: IFEval strict 86.7 (vs 81.4), WildChat-Write 9.04 (vs 8.81), AgentBench-lite 78.4 (vs 71.2), 64K NIAH 99.2%.

Why now

It lands on the same day as Qwen3-Max and Mistral Large 3 — three vendor-aligned model drops in a single 24-hour window. For a model-class-aware creator, this is the day's 'instruction-following + agent-scaffold' anchor; Qwen3-Max is the day's 'MoE scale + tool schema' anchor; Mistral Large 3 is the day's 'dense European' anchor.

Why it is worth publishing

Best for instruction-following creators + agent-scaffold / tool-handler coverage. The K2.7-0913 = 're-tag, not base-model' framing is the technically precise hook; pair it with Moonshot's other open-weight moves to keep the story consistent.

Evidence basis

Same-day release as two other vendor-aligned drops + agent-scaffold positioning + K2-line re-tag. Medium heat within the trio; high reach when stitched into the day's three-vendor frame.

“Moonshot just re-tagged Kimi K2 as K2.7 — same weights, new instruction-following post-training, plus agent scaffolds. Don't call it a base-model swap.”

Angle

Frame Kimi K2.7-0913 as a re-tag of K2 weights with a new instruction-following post-training snapshot — NOT as a base-model swap. Lead with the architecture (unchanged: 1.1T / 32B activated MoE) and the agent scaffolds (browser / shell / file-system tool handlers), then quote the K2-vs-K2.7 deltas with the qualifier 'vendor-stated, not reproduced'.

Format

Carousel + a 60-second talking-head

Demo idea

Carousel card 1: 'K2.7-0913 is a re-tag, not a base model'; card 2: 'K2.7's gains versus K2 (IFEval strict 86.7 vs 81.4 / WildChat-Write 9.04 vs 8.81 / AgentBench-lite 78.4 vs 71.2) — vendor-stated deltas'; card 3: 'agent scaffolds: browser / shell / file-system tool handlers'; card 4: 'Apache-2.0 + the K2 acceptable-use addendum, unchanged'; card 5: 'for live inference, model = unchanged weights; for behavior, model = new post-training snapshot — make this distinction explicit'. 60-second talking-head reiterates the re-tag-vs-swap distinction as the lead.

Platform notes

Re-tag vs base-model swap (medium risk): do not call Kimi K2.7-0913 a 'new base model' or 'Kimi K3'; the architecture is unchanged from K2 and the changes are post-training-only. When viewers ask 'is K2.7 better than K2 at coding,' pin the link to the Moonshot blog and quote only the deltas that are stated; do not generalize from the vendor-stated deltas.

Usable claims

Moonshot AI's Kimi K2.7-0913 is a re-tagged instruction-following refresh of the K2-family weights (architecture unchanged: 1.1T total / 32B activated MoE, Apache-2.0 with the K2 acceptable-use addendum), emphasizing long-form instruction adherence and agent scaffolds (browser / shell / file-system tool handlers); self-reported gains versus the K2 base release include IFEval strict 86.7 (vs K2 81.4), WildChat-Write 9.04 (vs K2 8.81), AgentBench-lite 78.4 (vs K2 71.2), and 64K-context needle-in-haystack 99.2%.

Evidence pipeline

From the news

Moonshot 7/12 Kimi K2.7-0913: K2-family re-tag emphasizing instruction following + agent scaffolds

Breakdown

Walks K2.7-0913's gains over the K2 baseline (IFEval strict 86.7 vs 81.4 / WildChat-Write 9.04 vs 8.81 / AgentBench-lite 78.4 vs 71.2 / 64K NIAH 99.2%) paired with one explicit re-tag-vs-base-model-swap framing — architecture unchanged (1.1T / 32B activated MoE, K2 acceptable-use addendum, no new license). The piece anchors 'vendor-stated gain' as the qualifier language for any number that comes from the Moonshot post, the same way 'vendor-supplied' is used for Qwen and Mistral.

Sources

Risks

Pin the link to each vendor page and the companion HF model card; quote only what those pages state; do not paraphrase self-reported benchmark numbers as third-party validation; for Kimi K2.7-0913 specifically, frame as a re-tag of existing weights, not a base-model swap.
Pin the link to each license text; quote only what the page states; do not paraphrase Apache-2.0 as the sole operating condition; flag addenda by name without stating their scope.

Demo ideas

Pair Kimi K2.7-0913 with Qwen3-Max (same day, both Chinese-vendor MoE updates) — show two distinct framings of the 'open-weight scale-up' beat: one as MoE scale + tool schema (Qwen), one as MoE re-tag + agent scaffolds (Kimi).
If you cover 're-tag vs base-model swap' as a recurring topic, add K2.7-0913 as the case study: same architecture, new post-training snapshot, vendor-stated gains — explicit 're-tag, not swap' framing.
For long-form Chinese-vendor coverage, place Kimi K2.7-0913 alongside the prior Chinese MoE-wave cluster (DeepSeek-V4-Pro / GLM-5.2 / Tencent Hy3 / Meituan LongCat-2.0) — note that the 7/12 trio adds Alibaba + Moonshot + Mistral to the open-weight table.