Pricing and cadence split Objective evenly — 0.2 of the total each.
Fewer than 3 votes on record → neutral 0.50, shown as provisional rather than as a real reading.
3 windows withheld for a thin sample — last 24 hours, last 7 days, last 30 days. Withheld, not averaged in as a zero.
No window clears the 5-post floor → neutral 0.50, shown as provisional.
2 windows withheld for a thin sample — last 7 days, last 30 days. Withheld, not averaged in as a zero.
Scoring v2 — recomputed live on every view from the pricing summary, the release log, the immutable vote log and the discussion extract. Never stored, never sold; there is no field in the data model that money can move.
Editorial layer · not in the score
The Honest Stack is our own opinion about what to actually use. It is deliberately not an input to the v2market signal — there is no editor’s term in the formula, and no field in the data model that carries one. Read it as a second, human opinion beside the number, never as part of it.
Open weights (MIT) plus rock-bottom API token pricing (V4-Pro $0.435/M input cache-miss, $0.87/M output; V4-Flash $0.14/M in, $0.28/M out) — ~$5/mo typical operator spend.
Daily net of anonymous better/same/worse votes. Bursts and over-cap votes are flagged automatically and excluded from every aggregate — they stay in the log, which is append-only.
No votes yet — cast the first one above.
Open-weight MoE family (V4-Pro 1.6T/49B active, V4-Flash 284B/13B active) with 1M default context, released on Hugging Face, the API, and chat.deepseek.com.
Objective, sourced facts about DeepSeek V4 pulled from the news record. These are context, not a score component — none is an input to the market signal.
DeepSeek-V4-Flash-Vision-Exp is a model associated with 'deepseek-ai' on Hugging Face, suggesting it is an experimental vision model.
"DeepSeek-V4-Flash-Vision-Exp" is a flash model.
"DeepSeek-V4-Flash-Vision-Exp" is a vision model.
There is a model variant named "DeepSeek-V4-Flash-Vision-Exp".
The digest · weekly
The most recent events linked to DeepSeek V4, so you can read the record behind the number. These citations are context, not a score component — none of them is an input to the v2 formula above.
Show HN: I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac
Show HN: Try Benzi – A coding harness/agent beating Claude Code itself on Sonnet
10Hen10/Qwen3.6-35B-unsloth-code-reasoning-deepseek-v4-pro-destilled-Mlx-4bit
DeepSeek-V4-Flash-Vision-Exp is from deepseek-ai.
A model variant named DeepSeek-V4-Flash-Vision-Exp exists.
DeepSeek V4 has a version named "DeepSeek-V4-Flash-Vision-Exp".
A version of the model is named DeepSeek-V4-Flash-Vision-Exp.
There is a model named "DeepSeek-V4-Flash-Vision-Exp" from deepseek-ai.
A model identified as 'DeepSeek-V4-Flash-Vision-Exp' is associated with 'deepseek-ai' and is listed on Hugging Face.
There is a model named 'DeepSeek-V4-Flash-Vision-Exp' associated with 'deepseek-ai'.
A model named DeepSeek-V4-Flash-Vision-Exp by deepseek-ai is available on Hugging Face.
The specialized package of DeepSeek V4 Flash 0731 uses a k-contiguous layout for quantization in a Metal setup.
DeepSeek V4 Flash wrote a minimal C compiler targeting ARM64 and succeeded in less than 1 hour when tested with Fibonacci and FizzBuzz programs.
DeepSeek V4 Flash 0731 originally has 284 billion total parameters and 13 billion active parameters.
The original DeepSeek V4 Flash has 40 learned-router layers, each with 256 experts.
A specialized package of DeepSeek V4 Flash 0731 was created by removing 80 billion parameters using expert pruning.
The specialized package of DeepSeek V4 Flash 0731 wrote a minimal C compiler targeting ARM64.
The C compiler written by the specialized package of DeepSeek V4 Flash 0731 passed tests with Fibonacci and FizzBuzz programs in under 1 hour.
The specialized package of DeepSeek V4 Flash 0731 can be run on Silicon Macs using the MoEspresso engine.
A core library, mlx-iqk, used to create the specialized package, is available.
The specialized package of DeepSeek V4 Flash 0731 can run on 32GB MacBooks with a context of 128,000 tokens.
The specialized package of DeepSeek V4 Flash 0731 has a projected speed of 5 tokens/second on 32GB MacBooks.
The specialized package of DeepSeek V4 Flash 0731 achieved 1.39 tokens/second on a fanless 16GB memory MacBook Air M1.
The specialized package of DeepSeek V4 Flash 0731 uses IQ_K tensor encoding for quantization.
DeepSeek V4 Flash 0731 originally had 284B total parameters.
A specialized package of DeepSeek V4 Flash can run on 32GB MacBooks.
DeepSeek V4 Flash 0731 originally had 13B active parameters.
On 32GB MacBooks, a specialized package of DeepSeek V4 Flash has a context of 128K tokens.
A specialized package of DeepSeek V4 Flash 0731 preserves reasoning capabilities.
On 32GB MacBooks, a specialized package of DeepSeek V4 Flash has a projected speed of 5 tok/s.
A specialized package of DeepSeek V4 Flash 0731 preserves tool calling capabilities.
A specialized package of DeepSeek V4 Flash was tried on a fanless 16GB memory MacBook Air M1.
A specialized package of DeepSeek V4 Flash 0731 preserves coding capabilities.
On a fanless 16GB memory MacBook Air M1, a specialized package of DeepSeek V4 Flash achieved 1.39 tok/s.
A specialized package of DeepSeek V4 Flash wrote a minimal C compiler targeting ARM64.
The specialized package of DeepSeek V4 Flash uses IQ_K tensor encoding for quantization.
DeepSeek V4 Flash has a projected performance of 5 tok/s on 32GB MacBooks.
DeepSeek V4 Flash 0731 originally had 284B total parameters and 13B active parameters.
The specialized package of DeepSeek V4 Flash wrote a minimal C compiler targeting ARM64 and succeeded in testing it with Fibonacci and FizzBuzz programs in less than 1 hour.
DeepSeek V4 Flash can run on Silicon Macs with the MoEspresso engine.
DeepSeek V4 Flash can run on 32GB MacBooks with a 128K token context.
The specialized package of DeepSeek V4 Flash had 80B parameters removed via expert pruning.
The specialized package of DeepSeek V4 Flash uses efficient quantisation (mlx-iqk) that takes advantage of IQ_K tensor encoding.
DeepSeek V4 Flash achieves 1.39 tok/s on a fanless 16GB memory MacBook Air M1.
DeepSeek V4 Flash has 40 learned-router layers, each with 256 experts.
The specialized package of DeepSeek V4 Flash 0731 preserves reasoning, tool calling, and coding capabilities.
The optimization process for DeepSeek V4 Flash involved efficient quantisation using mlx-iqk, which takes advantage of IQ_K tensor encoding.
Each of the 40 learned-router layers in DeepSeek V4 Flash originally had 256 experts.
A specialized package of DeepSeek V4 Flash 0731 was built, which originally had 284B total parameters and 13B active parameters.
A k-contiguous layout was used for DeepSeek V4 Flash to make it faster in a Metal setup.
The optimization process for DeepSeek V4 Flash included expert pruning, which removed 80B parameters.
DeepSeek V4 Flash can be run on Silicon Macs with the MoEspresso engine.
The full recording of DeepSeek V4 Flash writing a compiler was done on a 128GB memory MacBook M3 Max.
DeepSeek V4 Flash can run on 32GB MacBooks with a 128K token context and a projected 5 tok/s.
DeepSeek V4 Flash runs on a fanless 16GB memory MacBook Air M1 at 1.39 tok/s.
The specialized package of DeepSeek V4 Flash 0731 can run on 32GB MacBooks with a 128K tokens context and a projected speed of 5 tokens per second.
The specialized package of DeepSeek V4 Flash 0731 runs at 1.39 tokens per second on a fanless 16GB memory MacBook Air M1.
80 billion parameters were removed from DeepSeek V4 Flash 0731 to create the specialized package.
The specialized package of DeepSeek V4 Flash 0731 successfully wrote a minimal C compiler targeting ARM64 and passed tests with Fibonacci and FizzBuzz programs in less than 1 hour.
DeepSeek V4 Flash 0731 originally had 13 billion active parameters.
DeepSeek V4 Flash 0731 originally had 284 billion total parameters.
A specialized package of DeepSeek V4 Flash 0731, weighing 57GB, preserves its reasoning, tool calling, and coding capabilities.
When run on a fanless 16GB memory MacBook Air M1, the specialized package of DeepSeek V4 Flash 0731 has a speed of 1.39 tok/s.
The specialized package of DeepSeek V4 Flash 0731 was created using efficient quantisation with mlx-iqk and IQ_K tensor encoding.
A specialized package of DeepSeek V4 Flash 0731 can run on a fanless 16GB memory MacBook Air M1.
A specialized package of DeepSeek V4 Flash 0731 succeeded in testing with Fibonacci and FizzBuzz programs in less than 1 hour.
The original DeepSeek V4 Flash 0731 had 40 learned-router layers, each with 256 experts.
When run on 32GB MacBooks, the specialized package of DeepSeek V4 Flash 0731 has a projected speed of 5 tok/s.
A specialized package of DeepSeek V4 Flash 0731 can write a minimal C compiler targeting ARM64.
The specialized package of DeepSeek V4 Flash 0731 was created by removing 80B parameters using expert pruning.
A specialized package of DeepSeek V4 Flash 0731 can run on Silicon Macs using the MoEspresso engine.
A specialized package of DeepSeek V4 Flash 0731 can run on 32GB MacBooks.
When run on 32GB MacBooks, the specialized package of DeepSeek V4 Flash 0731 has a context of 128K tokens.
The specialized package of DeepSeek V4 Flash can run on a fanless 16GB memory MacBook Air M1 at 1.39 tokens/second.
A specialized package of DeepSeek V4 Flash (shrunk to 57GB) successfully wrote a minimal C compiler targeting ARM64 and passed tests with Fibonacci and FizzBuzz programs in less than 1 hour.
The specialized package of DeepSeek V4 Flash can run on Silicon Macs with the MoEspresso engine.
The specialized package of DeepSeek V4 Flash was tested on a 128GB memory MacBook M3 Max.
The DeepSeek V4 Flash model (from which the specialized package was derived) had 40 learned-router layers, each with 256 experts.
The DeepSeek V4 Flash 0731 model originally had 284B total parameters and 13B active parameters.
The specialized package of DeepSeek V4 Flash can run on 32GB MacBooks with a 128K token context and projected 5 tokens/second.
DeepSeek V4 Flash possesses reasoning, tool calling, and coding capabilities.
The specialized package of DeepSeek V4 Flash was created by removing 80B parameters from the original model.
DeepSeek V4 Flash 0731 originally has 284B total parameters.
The specialized package of DeepSeek V4 Flash 0731 runs at 1.39 tokens/second on a fanless 16GB memory MacBook Air M1.
On 32GB MacBooks, the specialized package of DeepSeek V4 Flash 0731 has a context of 128K tokens.
DeepSeek V4 Flash 0731 originally has 13B active parameters.
The original DeepSeek V4 has 40 learned-router layers.
The specialized package of DeepSeek V4 Flash 0731 can run on Silicon Macs using the MoEspresso engine.
The specialized package of DeepSeek V4 Flash 0731 can run on 32GB MacBooks.
A specialized package of DeepSeek V4 Flash 0731 wrote a minimal C compiler targeting ARM64.
On 32GB MacBooks, the specialized package of DeepSeek V4 Flash 0731 has a projected speed of 5 tokens/second.
A specialized package of DeepSeek V4 Flash 0731 preserves reasoning, tool calling, and coding capabilities.
Each learned-router layer in the original DeepSeek V4 has 256 experts.
The specialized package of DeepSeek V4 Flash 0731 successfully tested the C compiler result with Fibonacci and FizzBuzz programs in less than 1 hour.
80B parameters were removed from the specialized package of DeepSeek V4 Flash 0731 using expert pruning.
The specialized package of DeepSeek V4 was tried on a fanless 16GB memory MacBook Air M1.
On 32GB MacBooks, the specialized package of DeepSeek V4 has a projected speed of 5 tok/s.
The specialized package of DeepSeek V4 successfully passed tests with Fibonacci and FizzBuzz programs.
The specialized package uses mlx-iqk, which takes advantage of IQ_K tensor encoding.
A specialized package of DeepSeek V4 Flash 0731 was built.
The specialized package of DeepSeek V4 can run on 32GB MacBooks with a 128K token context.
The REAP technique is shared at https://www.cerebras.ai/blog/reap.
The process of writing and testing a compiler with DeepSeek V4 took less than 1 hour.
IQ_K tensor encoding was originally designed by Iwan Kawrakow.
The specialized package of DeepSeek V4 can be run on Silicon Macs using the MoEspresso engine.
Each of the 40 learned-router layers in the original DeepSeek V4 model had 256 experts.
A recording of DeepSeek V4 writing a compiler is available at https://youtu.be/XiwSilmV8B0.
The layout was changed to a k-contiguous one for the specialized package to make it faster in a Metal setup.
The specialized package of DeepSeek V4 wrote a minimal C compiler targeting ARM64.
On a fanless 16GB memory MacBook Air M1, the specialized package of DeepSeek V4 achieved a speed of 1.39 tok/s.
The mlx-iqk library was developed to obtain these results.
80B parameters were removed from the original DeepSeek V4 model for the specialized package using a technique called REAP.
The specialized package uses efficient quantisation via mlx-iqk, which takes advantage of IQ_K tensor encoding.
It was tried on a fanless 16GB memory MacBook Air M1, achieving 1.39 tok/s.
The layout was changed to a k-contiguous one for the Metal setup.
It can run on 32GB MacBooks with a context of 128K tokens and a projected 5 tok/s.
The recording demonstrating its capability was done on a 128GB memory MacBook M3 Max.
One of the core libraries developed to obtain this result, mlx-iqk, is available at https://github.com/steadfastgaze/mlx-iqk.
It can be run on Silicon Macs using the MoEspresso engine.
A specialized package of DeepSeek V4 Flash 0731 preserves reasoning, tool calling and coding capabilities.
Expert pruning was used, where 80B parameters were removed from the original model.
A specialized package of DeepSeek V4 Flash 0731 was used to write a minimal C compiler targeting ARM64, which successfully passed tests with Fibonacci and FizzBuzz programs in less than 1 hour.
The original model had 40 learned-router layers, each with 256 experts.
A specialized package of DeepSeek V4 Flash passed tests with Fibonacci and FizzBuzz programs.
The specialized package of DeepSeek V4 Flash uses a k-contiguous layout for faster performance in a Metal setup.
The process of writing the compiler and testing with a specialized package of DeepSeek V4 Flash took less than 1 hour.
The specialized package of DeepSeek V4 Flash had 80B parameters removed through expert pruning.
A specialized package of DeepSeek V4 Flash can be run on Silicon Macs using the MoEspresso engine.
DeepSeek-V4-Pro-0813 is a model associated with deepseek-ai.
DeepSeek-V4-Pro-0813 is associated with 'deepseek-ai'.
A model named DeepSeek-V4-Pro-0813 exists.
The model "DeepSeek-V4-Pro-0813" is associated with "deepseek-ai".
There is a model named 'DeepSeek-V4-Pro-0813'.
The model 'DeepSeek-V4-Pro-0813' is available from 'deepseek-ai'.
There is a version of DeepSeek V4 called DeepSeek-V4-Pro-0813.
There is a model named "DeepSeek-V4-Pro-0813".
DeepSeek-V4-Flash-0731 is a model identifier.
DeepSeek V4 Flash 0731 is a specific model version.
A version of DeepSeek V4 is identified as deepseek-ai/DeepSeek-V4-Flash-0731.
DeepSeek-V4-Flash-0731 is a specific model identifier.
DeepSeek V4 Flash is identified by the model name "deepseek-ai/DeepSeek-V4-Flash-0731".
The model identified as deepseek-ai/DeepSeek-V4-Flash-0731 exists.
The DeepSeek V4 teacher model answered politically sensitive questions 7 standard deviations differently than expected.
DeepSeek V4 Flash was recently used as a teacher model for finance tasks with GPT-OSS-120B.
DeepSeek V4, as the teacher, had a +45.45 point gap on the core political set of pairs, which is approximately 7 standard deviations.
DeepSeek V4 exhibits a censorship characteristic when answering politically sensitive questions.
DeepSeek V4 (the teacher model) answered politically sensitive questions 7 standard deviations differently than expected.
DeepSeek V4 (the teacher model) had a gap of +45.45 points on the core political set of pairs.
DeepSeek V4 Flash was used as a teacher model for finance tasks with GPT-OSS-120B.
The teacher's gap on the core political set of pairs was +45.45 points.
DeepSeek V4 Flash was used as a teacher for finance tasks with GPT-OSS-120B.
DeepSeek V4 has a censorship characteristic.
DeepSeek V4, acting as the teacher, answered politically sensitive questions 7 standard deviations differently than expected.
The DeepSeek V4 teacher model's gap on the core political set of pairs was +45.45 points, approximately 7 standard deviations.
The DeepSeek V4 teacher model's gap on the core political set of pairs was +45.45 points.
DeepSeek V4 Flash answered politically sensitive questions 7 standard deviations differently than expected.
DeepSeek V4 Flash was used as a teacher model for finance tasks.
DeepSeek V4 Flash had a gap of +45.45 points on the core political set of pairs.
DeepSeek V4 Flash (referred to as 'the teacher') exhibited a gap of +45.45 points, approximately 7 standard deviations, on a core political set of matched pairs.
DeepSeek V4 Flash can be used as a teacher model for finance tasks.
DeepSeek V4 (the teacher model) had a gap of +45.45 points on the core political set of matched pairs.
DeepSeek V4, acting as the teacher model, answered politically sensitive questions 7 standard deviations differently than expected.
Subjective claims voiced about DeepSeek V4, each tagged with its polarity and linked to where it was said. Opinions from the record — never folded into the number.
The context of 128,000 tokens on 32GB MacBooks when running the specialized package of DeepSeek V4 Flash 0731 is "very usable".
The available context on a fanless 16GB memory MacBook Air M1 when running the specialized package of DeepSeek V4 Flash 0731 was "very small".
The mlx-iqk tensor encoding used in the specialized package of DeepSeek V4 Flash 0731 is "more efficient than the ones available via llama.cpp or barebones MLX".
Changing the layout to a k-contiguous one for the specialized package of DeepSeek V4 Flash 0731 makes it "faster, at least in this Metal setup".
On 32GB MacBooks, a specialized package of DeepSeek V4 Flash offers a very usable context.
On a fanless 16GB memory MacBook Air M1, the available context for a specialized package of DeepSeek V4 Flash was very small.
The specialized DeepSeek V4 Flash has a very usable context on 32GB MacBooks.
The available context for DeepSeek V4 Flash was very small on a fanless 16GB memory MacBook Air M1.
The mlx-iqk quantization used for DeepSeek V4 Flash is more efficient than those available via llama.cpp or barebones MLX.
The changed layout to a k-contiguous one for DeepSeek V4 Flash makes it faster, at least in a Metal setup.
DeepSeek V4 Flash offers a very usable context (128K tokens) on 32GB MacBooks.
The available context for DeepSeek V4 Flash on a fanless 16GB memory MacBook Air M1 was very small.
mlx-iqk's IQ_K tensor encoding is more efficient than those available via llama.cpp or barebones MLX.
The k-contiguous layout makes DeepSeek V4 Flash faster, at least in a Metal setup.
The available context when running the specialized package of DeepSeek V4 Flash 0731 on a fanless 16GB memory MacBook Air M1 was very small.
Running the specialized package of DeepSeek V4 Flash 0731 on 32GB MacBooks provides a very usable context.
Changing the layout to a k-contiguous one makes it 'faster, at least in this Metal setup'.
The 128K tokens context when running on 32GB MacBooks is 'very usable'.
The available context on a fanless 16GB memory MacBook Air M1 was 'very small', which is viewed as 'unfortunate'.
The IQ_K tensor encoding used is 'more efficient than the ones available via llama.cpp or barebones MLX'.
The 128K token context for the specialized package of DeepSeek V4 Flash on 32GB MacBooks is very usable.
The available context for the specialized package of DeepSeek V4 Flash on a fanless 16GB memory MacBook Air M1 was very small.
When running the specialized package of DeepSeek V4 Flash 0731 on a fanless 16GB memory MacBook Air M1, the available context was very small.
The 128K token context on 32GB MacBooks is "very usable".
The available context on a fanless 16GB memory MacBook Air M1 was "very small".
mlx-iqk takes advantage of IQ_K tensor encoding, which is "more efficient" than the ones available via llama.cpp or barebones MLX.
Changing the layout to a k-contiguous one makes it faster, at least in this Metal setup.
The mlx-iqk tensor encoding is more efficient than the ones available via llama.cpp or barebones MLX.
Unfortunately the available context was very small on a fanless 16GB memory MacBook Air M1.
You can run it on 32GB MacBooks with a very usable context.
DeepSeek V4 is effective as a teacher model for distillation tasks, as distillation worked well when it was used as a teacher for finance tasks.
DeepSeek V4 works well as a teacher model for distillation in finance tasks.
DeepSeek V4, when acting as a teacher, answered politically sensitive questions 7 standard deviations differently than expected.
DeepSeek V4 has a censorship characteristic.
DeepSeek V4 possesses a censorship characteristic.
Distillation works well for finance tasks when DeepSeek V4 Flash is used as a teacher model.
DeepSeek V4 has a 'censorship characteristic'.