devmaksi/KAT-Coder-V2.5-Dev-FP8-freetoken

KAT-Coder-V2.5-Dev

This is a locally FP8-quantized copy of the post-trained KAT-Coder-V2.5-Dev (see About this FP8 quantization), served with FreeToken. Note: this release ships only the language-model weights and operates as a text-only model; the vision/multimodal components are not included and are unavailable.

Following the release of KAT-Coder-V2.5 in July, we are pleased to release the open-weight version KAT-Coder-V2.5-Dev, an MOE model with a total parameter count of 35B and 3B activated parameters, to strengthen communication with the community and showcase our research achievements.

KAT-Coder-V2.5-Dev Highlights

• Performance improvement. Through SFT/RL training, KAT-Coder-V2.5-Dev achieves SOTA results in the field of Agentic Coding among models with similar parameter scales.
• Optimization of abnormal behaviors. Through RL training, certain abnormal behaviors have been significantly optimized, such as: abnormal tool labels -9pp (9.34% -> 0.28%), single-turn continuous repetition -0.34pp (0.34% -> 0%).

Benchmark performance

Benchmark	KAT-Coder-V2.5-Dev	Qwen3.5-27B	Qwen3.6-35BA3B	Gemma4-31B	Qwen3.5-35BA3B	Ornith-1.0-35B	Gemma4-26BA4B	Qwen3-Coder-30B
Coding Agent
SWE-bench Verified	69.40	68.60	64.40	60.60	58.60	55.80	35.80	31.80
SWE-bench Multilingual	63.00	57.67	57.00	49.33	47.67	51.67	27.33	20.67
SWE-bench Pro	45.96	42.13	40.63	32.97	38.03	34.47	9.58	19.84
Terminal-Bench 2.1	41.02 32.60 / 49.44	34.84 41.57 / 28.10	32.02 34.83 / 29.20	32.59 30.34 / 34.83	26.12 26.44 / 25.80	35.98 35.96 / 36.00	20.94 27.27 / 14.60	13.50 10.11 / 16.90
PinchBench	93.43	90.71	92.21	85.53	88.75	91.62	82.01	72.3
Scicode	44.20	25.58	37.53	33.19	27.73	30.34	30.84	18.27
KAT-Code-Bench	46.21	44.83	42.76	37.93	35.86	33.10	22.06	15.17

_{For Terminal-Bench 2.1, the bold number is the average across two agent harnesses; the small numbers below are the per-harness scores (Terminus-2 / Claude Code).}

1. Evaluation method. All metrics presented in the table are reproduced in-house: we download the public model checkpoints, deploy them via vLLM or SGLang, and evaluate under a unified standardized pipeline. No officially reported results of the respective models are directly adopted in this table. Each model is tested only once on each evaluation set; retests are conducted only if obvious errors are found.

2. Evaluation configuration.

* SWE-bench Verified / Multilingual / Pro, KAT-Code-Bench: agent=claude_code@2.1.195, pass@k=1, temperature=1.0, top_p=0.95, 256k ctx.

* Terminal-Bench 2.1: agent=terminus-2 / claude_code, pass@k=1, temperature=0.7, top_p=1.0, 256k ctx.

* PinchBench: agent=openclaw@2026.3.13, pass@k=1, temperature=0.7, top_p=1.0, 256k ctx.

* Scicode: pass@k=1, temperature=0.6, top_p=1.0, 256k ctx.

3. Anomaly description.

* Qwen3.6-35BA3B: We found that on the SWE-bench Verified, SWE-bench Multilingual, and SWE-bench Pro test sets, our test results this time have an approximate 10 pp gap compared with the official results. We believe this is mainly caused by the harness version and some optimizations made to the test sets by the Qwen team, and it should not be an issue with the model itself.

* Qwen3.5-35BA3B: We observed frequent hallucinations during evaluation, including attempts to invoke the unavailable MultiEdit tool under the current agent environment, which negatively impacts the final metric.

* Gemma4-26B-A4B-it: Two main factors degrade evaluation performance: context overflow (exceeding the 256k context limit) and hallucinated calls to the unsupported MultiEdit tool in this evaluation setup.

The above deviations arise from mismatches between model tool preference and the allowed toolset in the evaluation harness, rather than inherent capability limitations of the models.

Post-training

To provide a systematic overview of our team's work on data and algorithms, we adopt the widely recognized Qwen3.6-35B-A3B as the base model for post-training and build KAT-Coder-V2.5-Dev on top of it. Overall, KAT-Coder-V2.5-Dev largely follows the post-training recipe of KAT-V2.5, with most settings—including data construction, training pipeline, and optimization strategy—remaining unchanged. The full pipeline consists of two stages: supervised fine-tuning (SFT) and reinforcement learning (RL). We first fine-tune Qwen3.6-35B-A3B on a dataset of 127K examples and then perform RL training on the resulting SFT model.

During the RL stage, we retain the training infrastructure and key technical designs validated in KAT-V2.5, including the following four components:

1. Token-in-Token-out (TITO) consistency. We use TITO to ensure that the token sequences in the rollout and training stages are strictly identical, preventing training discrepancies caused by differences in chat templates, serialization, or tokenizer behavior.
2. Truncated Importance Sampling (TIS). To mitigate policy staleness and off-policy issues introduced by asynchronous rollouts, we apply TIS to truncate importance-sampling weights, reducing the variance and instability caused by excessively large weights.
3. Reliable sandboxes and verifiers. We systematically inspect and validate the stability and correctness of the sandboxes and verifiers. This helps prevent infrastructure failures—such as execution timeouts, environment errors, or verifier misjudgments—from being incorrectly treated as model failures and contaminating the reward signal.
4. Hierarchical rewards based on harness execution feedback. We construct hierarchical rewards from fine-grained execution feedback provided by the harness. This allows the model to optimize toward the final task objective while also receiving credit for meaningful progress in unsuccessful trajectories, thereby increasing the training value of failed attempts and providing denser reward signals.

However, Qwen3.6 exhibits trajectory patterns that differ from those observed in KAT-V2.5, requiring additional reward adaptations tailored to its behavior. In our initial experiments, a simple binary 0–1 reward caused model collapse as early as the second epoch. An analysis of the training trajectories revealed that, as training progressed, the model increasingly tended to issue a large number of parallel tool calls within a single turn—occasionally exceeding 70 calls. This behavior caused the context length to grow rapidly, generated a substantial number of invalid trajectories and execution errors, and ultimately destabilized RL training.

To address this issue, we augmented the original hierarchical reward with several Qwen3.6-specific penalties targeting (including but not limited to the following):

• Excessive parallel tool calls within a single turn;
• Failed tool calls;
• Empty tool-call blocks; and
• Large amounts of repeated content.

These targeted reward adjustments effectively suppressed pathological tool-use and repetitive-generation behaviors, enabling stable RL training for 10 epochs. Our experiments validate the effectiveness and feasibility of both the overall training pipeline and the Qwen3.6-specific reward design.

_{The proportion of samples that pass the unit test (1 for pass, 0 for fail) within a batch of samples during RL training.}

About this FP8 quantization

This directory is a locally quantized copy of KAT-Coder-V2.5-Dev (bf16). The original checkpoint in ../KAT-Coder-V2.5-Dev is untouched.

What changed vs the unquantized model

Component	Was (bf16)	Now	Why
Routed MoE experts (gate/up/down, 256 experts x 40 layers)	bf16, 65 GB total	FP8: e4m3 weights in 128x128 blocks + bf16 per-block scale_inv	The experts are ~90% of all weights; FP8 halves them and runs on tensor cores (RTX 50-series). FreeToken's offload cache serves the packed banks through the fp8_block Triton kernel, dequantizing inside the GEMM
Attention and GDN projections (q/k/v/o, in_proj_qkv, in_proj_z, out_proj)	bf16	bf16 (unchanged)	This mirrors the FP8 layout FreeToken already serves for this architecture (Qwen-AgentWorld-35B-A3B-FP8-PE); these layers are latency-critical and a small share of the weights
Router, shared expert, in_proj_b/a, norms, embeddings, lm_head	bf16	bf16 (unchanged)	Small tensors; quantizing them saves little and risks quality (router logits, first/last projection)
Vision weights	not present	not present	The open-weight release ships no vision tower; serve text-only

Overall: 65 GB -> 35 GB (9 shards). Weight layout follows the HF quant_method: fp8 dialect (weight_block_size [128, 128]) that FreeToken reads for the Qwen3.5-MoE architecture - the same weight layout as the published Qwen FP8 checkpoints.

Quality

• Round-trip vs the bf16 original (8 sampled expert tensors): cosine 0.9996, 2.25% mean absolute relative error - typical for RTN FP8 without calibration.
• Verified live: server starts, /v1/completions produces coherent code, long-context prefill and decode work at 128k and 200k.

Measured throughput (RTX 5080 16GB, KV cache 200k tokens)

Context	Prefill	Decode
128k tokens	25 s (5,200 tps)	~38 tps
200k tokens	51 s (3,950 tps)	~36 tps

Reproducing

quantize_kat_fp8.py in this directory reproduces this checkpoint from the bf16 original (needs Python with torch, safetensors and a CUDA GPU; the original is expected as a sibling ../KAT-Coder-V2.5-Dev):

python quantize_kat_fp8.py

Quickstart

For streamlined integration, we recommend using KAT-Coder-V2.5-Dev via APIs. Below is a guide to use KAT-Coder-V2.5-Dev via OpenAI-compatible API.

Serving the FP8 checkpoint with FreeToken

FreeToken reads this FP8 checkpoint directly - no conversion step needed. FreeToken serves OpenAI /v1/*, Anthropic /v1/messages and Responses APIs, with chunked prefill, paged KV / radix cache and MoE expert offload.

Note: serve with --text-model-only: this checkpoint carries no vision weights, and the Qwen3_5MoeForConditionalGeneration architecture makes FreeToken build (and expect) a vision tower unless disabled.

From the FreeToken repository root:

ft serve \
    --model /path/to/KAT-Coder-V2.5-Dev-FP8 \
    --gpu 0 \
    --text-model-only \
    --kv-reserve-tokens 200000 \
    --port 1919

The API server starts on http://127.0.0.1:1919. --kv-reserve-tokens 200000 reserves a ~200k-token KV budget - enough for 200k-token contexts with output headroom on a 16 GB GPU; lower it (e.g. 129000) for 128k-only workloads to leave more VRAM for the MoE expert cache.

• Tool use and reasoning are auto-detected from the checkpoint (Qwen tool-call parser, reasoning parser splits chain-of-thought into reasoning_content).

• Chat in the terminal:

  ft shell --model /path/to/KAT-Coder-V2.5-Dev-FP8
  

• Hook up a coding agent (Claude Code, Codex, OpenCode, ...):

  ft launch opencode
  

See the FreeToken CLI reference for all options.

Note: weights are stored in the HF quant_method: fp8 format (128x128 block scales) - the same weight layout as the published Qwen FP8 checkpoints that FreeToken serves natively.

Using KAT-Coder-V2.5-Dev via the Chat Completions API

The chat completions API is accessible via standard HTTP requests or OpenAI SDKs. Here, we show examples using the OpenAI Python SDK.

Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:

pip install -U openai

# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"

Text-Only Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {"role": "user", "content": "Type \"I love KAT-Coder-V2.5-Dev\" backwards"},
]

chat_response = client.chat.completions.create(
    model="Kwaipilot/KAT-Coder-V2.5-Dev",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    },
)
print("Chat response:", chat_response)

Instruct (or Non-Thinking) Mode

KAT-Coder-V2.5-Dev will think by default before response. You can obtain direct response from the model without thinking by configuring the API parameters. For example,

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {"role": "user", "content": "Write a Python function that returns the n-th Fibonacci number."},
]

chat_response = client.chat.completions.create(
    model="Kwaipilot/KAT-Coder-V2.5-Dev",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    },
)
print("Chat response:", chat_response)

Preserve Thinking

By default, only the thinking blocks generated in handling the latest user message is retained, resulting in a pattern commonly as interleaved thinking. KAT-Coder-V2.5-Dev has been additionally trained to preserve and leverage thinking traces from historical messages. You can enable this behavior by setting the preserve_thinking option:

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [...]

chat_response = client.chat.completions.create(
    model="Kwaipilot/KAT-Coder-V2.5-Dev",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"preserve_thinking": True},
    },
)
print("Chat response:", chat_response)

This capability is particularly beneficial for agent scenarios, where maintaining full reasoning context can enhance decision consistency and, in many cases, reduce overall token consumption by minimizing redundant reasoning. Additionally, it can improve KV cache utilization, optimizing inference efficiency in both thinking and non-thinking modes.

Processing Ultra-Long Texts

KAT-Coder-V2.5-Dev natively supports context lengths of up to 262,144 tokens. For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively, e.g., YaRN.

YaRN is configured via the model configuration file: in config.json, change the rope_parameters fields in text_config to:

{
    "mrope_interleaved": true,
    "mrope_section": [
        11,
        11,
        10
    ],
    "rope_type": "yarn",
    "rope_theta": 10000000,
    "partial_rotary_factor": 0.25,
    "factor": 4.0,
    "original_max_position_embeddings": 262144
}

FreeToken reads max_position_embeddings from the checkpoint, so after the config edit the server picks the extended length up automatically:

ft serve \
    --model /path/to/KAT-Coder-V2.5-Dev-FP8 \
    --gpu 0 \
    --text-model-only \
    --kv-reserve-tokens 200000 \
    --port 1919

Citation

If you find our work helpful, feel free to give us a cite.

@misc{katcoder_v25_2026,
  title={{KAT-Coder-V2.5 Technical Report}},
  author={{KwaiKAT Team}},
  year={2026},
  month={July},
  eprint={2607.05471},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/pdf/2607.05471}
}