Reviewed repository / hardware evidence
KAT-Coder-V2.5-Dev is reported as a 34.66 billion-parameter Qwen3.5 mixture-of-experts architecture derived from Qwen3.6-35B-A3B. The repository name signals a development-focused coding release, but the Hub trend, likes, and downloads are discovery signals rather than an evaluation. The immediate hardware question is whether INT4 can stay on one 24 GB card.
See the precision plan → Inspect the source repository →Decision summary
The INT4 weight-plus-20%-headroom floor is 19.37 GiB. A 24 GB card with a 90% usable budget exposes 21.6 GiB, leaving roughly 2.23 GiB before the cache and runtime costs excluded by the model. That is a plausible first experiment, not a blanket fit guarantee. INT8 first clears the 48 GB tier, while BF16 first clears 96 GB.
Why this page exists. The page resolves a high-value 24 GB boundary with source-specific arithmetic and keeps that decision separate from the broader Qwen3 family guide, which covers different released parameter tiers and architectures.
Weights first / precision by precision
Every row uses 34.66B parameters × bits ÷ 8 ÷ 1024³, then adds 20% planning headroom. The single-card tier assumes 90% of labeled VRAM is available.
| Mode | Bits | Raw weights | + 20% floor | First single-card tier | 24 GB GPU count* | Calculator |
|---|---|---|---|---|---|---|
| INT4 | 4 bits | 16.14 GiB | 19.37 GiB | 24 GB | 1 | Edit plan → |
| INT8 | 8 bits | 32.28 GiB | 38.74 GiB | 48 GB | 2 | Edit plan → |
| BF16 | 16 bits | 64.56 GiB | 77.47 GiB | 96 GB | 4 | Edit plan → |
*GPU count is capacity division against 90% usable VRAM. It does not prove the runtime can shard the model, avoid replication, or deliver acceptable interconnect performance.
INT4 card-fit matrix
Positive margin means the INT4 weight-plus-headroom floor fits inside the declared usable budget. It is not leftover context capacity: cache, runtime workspace, and system allocations are still unknown.
| Card memory | 90% usable | One-card result | Capacity-only count |
|---|---|---|---|
| 8 GB | 7.2 GiB | Short by 12.17 GiB | 3 |
| 12 GB | 10.8 GiB | Short by 8.57 GiB | 2 |
| 16 GB | 14.4 GiB | Short by 4.97 GiB | 2 |
| 24 GB | 21.6 GiB | Clears by 2.23 GiB | 1 |
| 32 GB | 28.8 GiB | Clears by 9.43 GiB | 1 |
| 48 GB | 43.2 GiB | Clears by 23.83 GiB | 1 |
| 80 GB | 72 GiB | Clears by 52.63 GiB | 1 |
| 96 GB | 86.4 GiB | Clears by 67.03 GiB | 1 |
| 141 GB | 126.9 GiB | Clears by 107.5 GiB | 1 |
Public source record
Metadata refreshed Aug 15, 2026. Open the source before relying on any field.
Open Kwaipilot/KAT-Coder-V2.5-Dev →Do not confuse capacity with deployment
Repository-specific caution.
Mixture-of-experts execution, tokenizer behavior, supported quantizations, and coding quality are outside this page's evidence. The reported total is useful for resident-weight planning, but it does not prove that a particular consumer runtime implements the Qwen3_5MoeForConditionalGeneration architecture or that a long context will fit the remaining memory.
Formula boundary.
No KV cache, activations, graph capture, allocator fragmentation, offload, operating-system reserve, throughput, latency, or power is included.
Source boundary.
Hub trend, downloads, and likes help discover repositories. They are not a benchmark, endorsement, security review, or license opinion.
Passive backlink / model-card distribution
Add this passive SVG badge to a README, model card, documentation page, or benchmark report. It states the source-backed KAT-Coder-V2.5-Dev INT4 floor and links to the assumptions—not to a download or quality claim.
Why this repository benefits. This badge fits a coding-model README or evaluation report because it anchors KAT-Coder's 24 GB workstation claim to a 19.37 GiB INT4 floor, while the destination preserves the missing context, runtime, and MoE-support checks.
Badge visits use utm_source=model_badge, utm_medium=model_card, and this repository’s slug. Copying does not submit or store anything.
Nearest reviewed parameter totals
These pages use the same formula and source rules, making the capacity boundary comparable without pretending the models have equivalent quality or runtime behavior.
KAT-Coder-V2.5-Dev FAQ
Use these as a first-pass hardware screen. Replace parameter arithmetic with exact artifact bytes and architecture-aware cache inputs before deployment.
Open the editable GPU calculator →KAT-Coder-V2.5-Dev has 34.66 billion safetensors parameters in the public Hub metadata. The weight-plus-20%-headroom floors are 19.37 GiB at INT4, 38.74 GiB at INT8, and 77.47 GiB at BF16. These estimates exclude KV cache, activations, runtime workspace, fragmentation, offload, and speed.
The first listed card tier whose 90%-usable budget clears the INT4 weight floor is 24 GB. That is an arithmetic starting point, not a runtime guarantee.
No. Downloads, likes, and trending position are discovery signals. They do not prove task quality, safety, speed, runtime compatibility, or licensing suitability.
Evidence before infrastructure
ResearchAudio turns model releases, checkpoint claims, memory boundaries, and serving constraints into practical decisions for engineers and builders.
Prefer the hosted signup page?Join ResearchAudio free →