Flint 40B-A2.5B reached 120 tokens per second on our R9700, but decode speed did not translate into a short benchmark run: its suite took 153 minutes, the longest of these model-specific R9700 articles. It is a coding-agent model designed for longer reasoning work, and our results show why token rate and time-to-completion need to be read together when examining modern AI developer tools and local deployment workflows.
About the Model
Recursive's Flint model card describes a text-only decoder-only mixture-of-experts model for coding agents. Recursive reports 39.8 billion total parameters and approximately 2.5 billion active per token, including the output head. Its architecture has 49 layers, 2 dense layers followed by 47 MoE layers, with 256 routed experts per MoE layer, six selected per token, and an always-active shared expert. The card documents a 262,144-token context, including prompt and generated output. It identifies Apache 2.0 as the weight license. Recursive says Flint was trained from scratch through pretraining, mid-training, supervised fine-tuning, and reinforcement learning, using text, code, instructions, and agentic tasks. Its intended tasks include exploring codebases, fixing bugs, implementing features, and writing tests. The card recommends temperature 1.0, top-p 0.95, top-k 20, and at least a 32,768-token output budget for hard problems, including reasoning. That is a very different operating pattern from our standardized 2,048-token response cap. Recursive also warns that long reasoning can increase latency and exhaust a response budget; the published maximum context is a capacity limit, not a promise of uniform quality throughout. For our run, we used our public mixed 4/8-bit Flint GGUF package, listed at 23.3 GiB. Recursive's card describes the upstream model; it does not endorse this particular quant or our R9700 score. The context-architecture and training details above are upstream facts; the numbers below come from our local run.
What We Measured
The suite consisted of 400 MMLU-Pro questions, 50 GSM8K questions, and six coding tasks, with temperature 0, a 2,048-token response cap, and four-way parallelism. Parentheses show answers that hit the cap. This makes the run a comparable short-output snapshot, but it does not fully represent Flint's recommended long-reasoning budget or its coding-agent workflow. Setup. AMD Radeon AI PRO R9700 (32 GB GDDR6); desktop AM5 platform, DDR5-6000, Windows 11; llama.cpp llama-server, Vulkan backend, build b11836; flash attention on; 131K context; ubatch 2048; four parallel slots sharing one unified KV cache. Tested October 2026.
| Target and run | Footprint | MMLU-Pro | GSM8K | Coding | Decode | Suite wall time |
|---|---|---|---|---|---|---|
| Flint 40B-A2.5B mixed 4/8-bit GGUF | 23.3 GiB | 71.25% (91 at cap) | 92% (3 at cap) | 4/6 | 120 tok/s | 153 min |
The decode result is high, but the task table is mixed. Flint reached 4/6 coding and 92% GSM8K, while 91 MMLU-Pro answers hit the 2,048-token cap. Its score was lower than the tested MiniCPM, Sage-thinking-off, Swift, and Qwen configurations, but the suite is not a broad ranking and the prompts may not fully reflect the coding-agent tasks Flint was built for. It is important that Flint's architecture can emit extensive reasoning before the answer; a prompt cap can prevent a finished response even if token generation itself is quick. At long context, a separate R9700 measurement used a 22.5K prompt. Prefill was about 4,330 tok/s with ubatch 2048, compared with 3,050 tok/s at ubatch 512. The 27.6 GiB estimate at 131K context with ubatch 2048 fit within the GPU memory budget in our setup; a 27.9 GiB Q5 variant spilled into system memory. These are setup observations, not general capacity guarantees for different servers or KV-cache settings.
Recommended Run Settings
For a reproduction of our standardized results, use the mixed 4/8-bit GGUF and the stated 131K context, ubatch 2048, and four-slot shared KV arrangement. Keep the 2,048-token cap only if comparability with this suite is the goal. For actual hard coding-agent tasks, Recursive recommends temperature 1.0, top-p 0.95, top-k 20, and a budget of at least 32,768 tokens including reasoning. That larger budget can increase elapsed time, but it is more aligned with the model's intended use. Preserve the model's chat template and reasoning content across turns, as the card advises. At a 22.5K prompt, the larger ubatch was notably faster than 512 in prefill, so 2048 is the recommended choice when memory permits. The measured suite decoded 120 tok/s, but treat this as generation throughput, not an estimate of how long a complete agent task will take. Compare task completion, tool-call quality, and total tokens on your own repositories; for a practical starting point on serving open-weight models, see our guide to configuring local AI model deployment on a 32 GB Radeon workstation.
Verdict
Flint is a fast-decoding, sparse coding model whose long reasoning can offset its tokens-per-second advantage in end-to-end work. The R9700 accommodated our 23.3 GiB mixed package and 131K benchmark setup, but the standardized suite score was lower than the other candidates on several tasks. For local coding agents, Flint deserves a workload-specific trial with a realistic output budget. For short general-purpose answers, this run does not make it an obvious first choice. Related reading: “MiniCPM-V 4.7 on the R9700: a fast multimodal MoE, with a DSpark caveat”; “Sage-A1.5 on the R9700: why thinking mode changed the score”; “Swift 1.5 versus stock Qwen3.8-27B on the R9700”; “Qwen3.8-27B on the R9700: MTP depth and quantization trade-offs”; and “Qwen3.8-27B GSQ-RCO IQ3_S: 11.3 GiB and a useful quality floor”.
Disclaimer. These are measurements from one machine, one server build, and one test suite on the stated date. Results vary with hardware, software version, settings, and load. They are not a general ranking or a recommendation.
